GBClaw: An agentic AI framework for atomistic grain boundary research and mechanistic discovery

1

GBClaw: An agentic AI framework for atomistic grain boundary research and

mechanistic discovery

Yazhuo Liua,† , Kunqing Dinga,† , Xiaoguo Gongb , Zifeng Wang a , Yichen Yang a , Yifeng Chea , Wen

Chend,e , Josh Kacherc , Olivier Pierron a , Shigenobu Ogatab , Ting Zhua,c, *

a Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA,

30332, USA

b Department of Mechanical Science and Bioengineering, Graduate School of Engineering

Science, The University of Osaka, Osaka, 560-8531, Japan

c School of Materials Science and Engineering, Georgia Institute of Technology, Atlanta, GA,

30332, USA

d Department of Aerospace and Mechanical Engineering, University of Southern California, Los

Angeles, CA, USA

e Mork Family Department of Chemical Engineering and Materials Science, University of

Southern California, Los Angeles, CA, USA

2

Abstract

Agentic artificial intelligence (AI) systems hold immense potential for automating complex

workflows in materials discovery and design, but extending them to domain-intensive mechanistic

investigations remains challenging. Here, we present GBClaw, an agentic AI framework for

investigating the structure and deformation of general grain boundaries (GBs). Although general

GBs constitute the vast majority of real-world interfaces in polycrystalline materials, they remain

understudied because their structural complexity and interfacial mismatch complicate

experimental and modeling research. GBClaw integrates large language models (LLMs) into a

modular architecture comprising an orchestrator, specialized agents, domain-specific tools,

simulation interfaces, and hierarchical memory. Through dynamic replanning and iterative

workflow refinement, the framework effectively connects systematic simulations with targeted

mechanistic analysis. Applied to 14 general asymmetric tilt GBs in nickel, GBClaw characterized

interfacial defect structures and identified diverse deformation modes across these

incommensurate boundaries. Comparing boundary sliding with and without secondary GB misfit

dislocations revealed substantial differences in shear strength, underscoring the critical role of

defect hierarchies in sliding resistance. Furthermore, researcher-guided follow-up calculations

traced an anomalously high shear resistance to metastable boundary structures. This work

demonstrates that combining agentic AI with domain expertise bridges high-throughput

computational screening with deep physical understanding of complex material interfaces.

† These authors contributed equally to this work.

* Corresponding author: ting.zhu@me.gatech.edu

3

Introduction

Agentic artificial intelligence (AI) systems offer new ways to automate workflows and support

discovery 1,2 . These systems integrate large language models (LLMs) with a harness comprising

prompts, tools, middleware, and other components that enable iterative reasoning and task

execution. Multiple agents can coordinate specialized capabilities to carry out complex workflows

involving many interdependent steps. Although applications in materials research increasingly

encompass information retrieval, computation, and data analysis 3,4 , extending these capabilities to

open-ended mechanistic investigations remains challenging. Such investigations require domain

expertise to define research objectives, assess physical validity, resolve ambiguities, and select

informative follow-up calculations 5,6 . Integrating researchers into the agentic workflow provides

an effective means for addressing these requirements 5,7 .

Grain boundary (GB) research presents a demanding application for this approach. GBs

separate adjacent grains in polycrystalline materials and fundamentally dictate bulk mechanical

behavior, including yield strength, tensile ductility, creep resistance, and fracture toughness 8–10 .

Although general GBs constitute the vast majority of real-world interfaces, they remain

understudied because their structural complexity and interfacial mismatch complicate atomistic

modeling. The macroscopic geometry of a planar GB possesses five degrees of freedom, and each

geometry can admit multiple metastable atomic configurations 11–13 . Characterizing this structural

diversity and its mechanical consequences 14,15 requires both consistent computational procedures

across many interfacial configurations and adaptive investigations guided by intermediate results.

A master agent (i.e., an orchestrator) can coordinate these activities by translating research

objectives into tasks for specialized agents that use validated tools, generate custom code, or

interface with scientific software6,16 . Domain experts contribute by evaluating the results and

identifying physical questions that require further investigation.

Here, we introduce GBClaw, an agentic framework built on the open-source LangChain

platform 17 for automated GB simulations with researcher-guided mechanistic analysis. We

demonstrate its capabilities on general asymmetric tilt GBs that lack an exact coincidence site

lattice18 (i.e., non-CSL). GBClaw characterizes the atomic structures, shear strengths, and

deformation pathways of these incommensurate GBs 12,19 , while employing automated procedures

to identify interfacial misfit defects. The framework flags anomalous mechanical responses and

dynamically refines workflows in response to researcher queries, conducting targeted simulations

4

that uncover underlying physical mechanisms. These applications demonstrate how a human-in-

the-loop agentic framework bridges systematic computational screening with a deep physical

understanding of complex material interfaces.

GBClaw: Architecture and workflow

Architecture

GB investigations combine sequential dependence with opportunities for parallel execution.

Downstream tasks often depend on earlier results, whereas independent cases within a stage can

be processed concurrently. Coordinating these activities requires task planning, information

exchange, and persistent tracking of the investigation state. Leveraging the strong reasoning and

analytical capabilities of LLMs, GBClaw supports these functions through a modular architecture

consisting of an orchestrator, specialized agents, domain-specific tools, computational interfaces,

and a file-based hierarchical memory system (Fig. 1). Researchers interact directly with the

orchestrator, supported by the planning agent, to define research objectives, review intermediate

findings, and refine subsequent investigations. The orchestrator delegates tasks, integrates the

results, and produces a research report. Each specialized agent is routed to an LLM through an

application programming interface (API) and operates with a bounded context, tailored prompt

instructions, and designated tools. Separating coordination from execution limits the context

required by each agent and supports concurrent processing of independent cases.

GBClaw includes five specialized agents with complementary roles (Fig. 1). The planning

agent uses the scientific knowledge encoded in LLMs 7 to translate research objectives into specific

questions, simulation parameters, and executable plans. The structure agent determines grain

orientations, symmetry-consistent pairs of terminating planes, and structural periodicities needed

for bicrystal construction. It supports both exact CSL geometries 20 and near-CSL constructions

that accommodate departures from exact coincidence21 . The LAMMPS agent constructs bicrystals

and performs structural relaxation, GB energy calculation, shear simulations, and minimum-

energy-path (MEP) searches using the nudged elastic band (NEB) method 22,23 . The analysis agent

links shear stress responses to atomic displacements, structural changes, and GB transition

pathways. It also identifies stress-drop events and generates visualizations to support mechanistic

interpretation. The coding agent generates, executes and refines custom scripts for computational

and data-processing tasks not covered by predefined tools 6,24 . The agents access LAMMPS 25,

OVITO26 , and ASE27 through a unified Python integration layer. This modular design supports the

5

addition of agents with further scientific capabilities.

Fig. 1 | Agentic architecture of GBClaw for researcher-guided grain boundary research. The

orchestrator serves as the central coordination hub, interfacing the researcher with five specialized

agents, domain-specific scientific software tools, and a file-based hierarchical memory system.

The modular architecture enables automated task decomposition and delegation, parallel

simulation execution, structured information exchange, and interactive, human-in-the-loop

workflow refinement.

GBClaw maintains continuity across tasks through structured information exchange. Each

delegated task receives the context for execution, including established findings and references to

previous simulations and computational artifacts. Artifacts include generated files, datasets,

simulation outputs, code, and analysis results. When additional information is needed, specialized

agents retrieve relevant records from project memory before invoking computational tools. On

completion, each agent reports its status, findings, and artifact references to the orchestrator, which

updates the project state and uses these results to guide subsequent tasks and researcher review.

6

This approach provides focused task context while preserving access to the broader investigation.

Because scientific interpretation depends on the model, computational conditions, and analysis

procedure as well as final output, GBClaw retains these through hierarchical memory and context

management as discussed next.

GBClaw employs hierarchical memory. Inspired by the CoALA framework of working,

episodic, semantic and procedural memory 28 , GBClaw organizes information into five agent-

facing levels (Supplementary Table S1). (i) Domain knowledge encodes workflows and task rules

within system instructions and tools. (ii) The project summary tracks top-level objectives, key

findings and conclusions for ongoing coordination. (iii) The conversation state records recent agent

exchanges, decisions and open tasks. (iv) Persistent research records store durable findings, run

summaries, and output references. (v) The artifact level contains underlying configurations,

trajectories, logs, code and analysis outputs. Beyond these agent-facing levels, GBClaw maintains

a separate archive containing complete interaction and tool execution histories for researcher

review. Excluding this archive from routine agent retrieval reduces the risk of treating unverified

conversational content as validated findings, while retaining records of how the investigation

developed.

GBClaw manages active context through summary updates and conversation compaction.

Long context can impair an LLM’s retrieval and use of relevant information29 , so the system

maintains concise project summaries and periodically condenses conversation histories. Before

proposed research records are stored, provenance checks link their claims to supporting sources

and simulation results. These checks support traceability but do not independently establish

physical correctness. Project summaries incorporate new findings and simulation records, while

conversation compression retains key decisions, pending tasks, and artifact references without

deleting the underlying computational data. When more detail is needed, agents use shared

identifiers to retrieve indexed records and exact run files. These persistent links also allow

downstream simulation and analysis tools to access the original datasets directly (Supplementary

Fig. S1). Such management connects scientific interpretation to its supporting models and outputs,

thereby enabling researchers to review evidence and guide further investigation.

Structural screening and targeted simulations

General asymmetric, non-CSL tilt GBs are common in polycrystalline materials, yet they

remain insufficiently understood. These boundaries provide an ideal platform for investigating

how complex interfacial structures govern inelastic deformation. Experimental studies have

7

demonstrated that GBs frequently feature low-index terminating planes on one or both sides 19,30–

36 , motivating our focus on GBs bounded by simple crystallographic terminations. In these non-

CSL tilt GBs, the opposing surface lattices exhibit an irrational ratio of repeat distances along the

sliding direction, rendering them inherently incommensurate19,34 . This interfacial mismatch can be

accommodated either by elastic strain or through a hierarchy of GB misfit dislocations in extended

boundaries 11,33,37 . Under shear loading, these interfaces can undergo pure sliding (tangential

translation only) or shear-coupled migration (tangential translation paired with normal boundary

motion)14,38 . Resolving these deformation mechanisms requires both high-throughput structural

screening and detailed analysis of local atomic rearrangements during deformation.

Fig. 2 | Agentic GBClaw workflow for systematic screening and mechanistic analysis of GB

shear strength. The workflow combines automated exploration with human-in-the-loop domain

guidance across six key stages: crystallographic screening, simulation planning, parallel bicrystal

construction and structural analysis, researcher review, dynamic replanning, and targeted

investigation of anomalous shear responses in general asymmetric ⟨110⟩ tilt GBs.

GBClaw investigated flat, atomically sharp tilt GBs in face-centered cubic (FCC) Ni bicrystals

through systematic structural screening combined with researcher-guided shear simulations and

analysis (Fig. 2). ⟨110⟩ tilt GBs were selected because they are frequently observed in polycrystals,

including in HRTEM images 35,36 , thereby providing structural benchmarks for evaluating atomistic

8

models. The detailed workflow is presented in Supplementary Fig. S2. Representative researcher-

GBClaw interactions are shown in Supplementary Fig. S3 and the accompanying demonstration

page (https://zhugroup.gatech.edu/gbclaw-demo). During screening, the structure agent used an

LLM to identify near-Σ GBs containing low-index terminating planes and having misorientations

close to those of reference Σ GBs 39,40 . Candidate boundaries were required to satisfy three criteria:

(i) at least one terminating plane belongs to the {100}, {110}, or {111} family; (ii) the deviation

of misorientation angle from the corresponding reference Σ GB does not exceed 5°; and (iii) the

absolute values of all Miller indices of both terminating planes are less than 4. The selected

boundaries were asymmetric and incommensurate. Their bicrystallographic parameters were then

used to construct the corresponding bicrystal models, which the LAMMPS agent relaxed to

characterize their interfacial defect structures.

Relaxation of extended boundaries produced a hierarchy of GB misfit dislocations 21 . Closely

spaced, non-glissile primary GB misfit dislocations (PGMDs) accommodate local mismatch,

whereas more widely spaced glissile secondary GB misfit dislocations (SGMDs) accommodate

the residual mismatch. The shortest complete repeat of the PGMD arrangement defines the primary

periodicity λ, while the characteristic spacing between SGMDs defines the secondary periodicity

Λ within a zeroth-order approximation21 . General procedures are developed to identify SGMDs

and analyze their characters. Although SGMD formation is usually energetically favorable in

extended boundaries, models that exclude these defects accommodate the residual mismatch

through elastic strain. The following sections present the screening results, compare the shear

responses with and without SGMDs, and describe the researcher-guided investigation of an

anomalous shear response.

Results

GB sliding and shear strength

To investigate the shear behavior of asymmetric non-CSL GBs, the structure agent applied the

screen criteria described above to general asymmetric ⟨110⟩ tilt GBs near reference Σ GBs. This

screening yielded 12 distinct pairs of terminating planes, each corresponding to an asymmetric

non-CSL GB (Table 1). Some plane pairs were associated with multiple reference Σ GBs. For

example, the {111}/{110} pair lies near both the Σ9 and Σ27 GBs. A given plane pairs can also

support multiple crystallographically distinct misorientations. For example, the {111}/{331} pair

includes variants at approximately 22° and 49°, whereas the {111}/{221} pair includes variants at

9

approximately 16° and 55°. In addition, the initial screening near Σ33a identified only the 22° and

16° configurations. To enable direct comparisons between the variants, researcher guidance added

the corresponding 49° and 55° configurations, increasing the final set to 14 distinct GB

configurations.

Table 1 | Reference Σ GBs with ⟨110⟩ tilt axis and low-index near-Σ GBs for Σ < 50. STGB:

symmetric tilt GB; ATGB: asymmetric tilt GB. θΣ and θ denote the misorientation angles of Σ

and near-Σ GBs, respectively; Δθ denotes their difference.

Exact Σ GBs Near-Σ GBs

Σ θΣ (°) GB plane 1 GB plane 2 Character Approx. plane 2 Δθ (°) θ (°) GB plane pair

3

70.53 (1 -1 1) (1 -1 -1) STGB

70.53 (0 0 -1) (-2 2 -1) ATGB

70.53 (-1 1 0) (-1 1 4) ATGB

70.53 (1 -1 -1) (-1 1 -5) ATGB

9

38.94 (-1 1 -1) (-1 1 -5) ATGB

38.94 (1 -1 0) (7 -7 8) ATGB (1 −1 1) 3.68 35.26 {111}/{110} 38.94 (-1 1 -1) (-11 11 1) ATGB (1 −1 0) 3.68 35.26

38.94 (0 0 -1) (4 -4 -7) ATGB

(1 −1 −2) 3.68 35.26 {100}/{112}

(2 −2 −3) 4.37 43.31 {100}/{223}

11

50.48 (0 0 -1) (-6 6 -7) ATGB (−1 1 −1) 4.26 54.74 {111}/{100} 50.48 (-1 1 -1) (-1 1 -19) ATGB (0 0 -1) 4.26 54.74

50.48 (1 -1 0) (7 -7 12) ATGB

(1 -1 2) 4.26 54.74 {110}/{112}

(2 −2 3) 3.79 46.69 {110}/{223}

17

86.63 (0 0 -1) (12 -12 -1) ATGB (1 −1 0) 3.37 90.00 {100}/{110} 86.63 (1 -1 0) (1 -1 24) ATGB (0 0 1) 3.37 90.00

19

26.53 (-1 1 0) (-17 17 12) ATGB (−3 3 2) 1.29 25.24 {110}/{332}

26.53 (0 0 -1) (-6 6 -17) ATGB (−1 1 −3) 1.29 25.24 {100}/{113}

27

31.59 (1 -1 -1) (11 -11 -1) ATGB (1 −1 0) 3.68 35.26 {111}/{110}

31.59 (0 0 1) (10 -10 23) ATGB (1 −1 2) 3.68 35.26 {100}/{112}

33a 20.05 (1 -1 -1) (13 -13 -5) ATGB

(3 −3 −1) 1.95 22.00 {111}/{331}

(2 −2 −1) 4.26 15.79 {111}/{221}

33c

58.99 (-1 1 -1) (1 -1 -19) ATGB (0 0 -1) 4.26 54.74 {111}/{100} 58.99 (0 0 -1) (-20 20 -17) ATGB (-1 1 -1) 4.26 54.74

43 80.63 (0 0 -1) (30 -30 -7) ATGB (3 −3 −1) 3.89 76.74 {100}/{331}

Following selection of the asymmetric non-CSL GBs, the LAMMPS agent constructed and

relaxed the corresponding Ni bicrystals, which were subsequently analyzed by the analysis agent

to evaluate their interfacial atomic configurations. Interatomic interactions were modeled using

10

the angular-dependent potential developed by Howells and Mishin41 , which reproduced the

experimentally observed interfacial motif of the {111}/{331}(22°) GB 35 . This boundary

exemplifies the hierarchical formation of PGMDs and SGMDs. Energy relaxation of an extended

boundary without PBCs along the horizontal direction produced closely spaced PGMDs with a

Burgers vector of bp = 1/2⟨110⟩{111} alongside more widely spaced SGMDs. As shown in Fig.

3a, b, small and large red regions with high atomic volumetric strains near the {331} terminating

plane identify the cores of PGMDs and SGMDs, respectively. The primary periodicity, defined by

the repeating arrangement of PGMDs, is λ = 5/2〈112〉{111}, referenced to the grain terminated

by the {111} plane. The secondary periodicity 𝚲𝚲 is the spacing between adjacent SGMDs. The

magnitude of |𝛌𝛌| = 2.16 nm is substantially smaller than that of |𝚲𝚲| = 15.73 nm (Fig. 3b).

To distinguish macroscopically uniform GB shearing from sliding mediated by SGMD glide,

the LAMMPS agent constructed a compact periodic bicrystal model based on the primary

periodicity λ. An interfacial region between neighboring SGMDs in the extended GB (Fig. 3c) was

isolated to build a supercell with PBCs within the GB plane (Fig. 3d). After energy minimization,

the supercell retained its local PGMD arrangement, while accommodating residual interfacial

mismatch through elastic strain (Fig. 3c, d). These compact models enabled the evaluation of

athermal shear strengths and deformation modes in the absence of SGMDs under imposed periodic

constraints.

Using these compact models, the orchestrator coordinated multiple LAMMPS agents to

parallelize shear simulations across all 14 GB configurations in opposing shearing directions (Fig.

3e). Horizontal shear displacement was applied incrementally, followed by energy minimization

after each increment. The shear strength 𝜏𝜏 0, defined as the peak shear stress in the subsequent steady response, varied widely among these boundaries, ranging from tens of MPa to 9.3 GPa.

Exceptionally low 𝜏𝜏 0 values of 23 MPa and 14 MPa for the {111}/{110} and {111}/{100} GBs suggest potential for low-friction sliding. In contrast, the high 𝜏𝜏 0 value of 3.5 GPa for the {100}/{110} GB challenges its previously proposed suitability as a low-friction interface42 .

Pronounced stress drops accompanied diverse inelastic deformation modes, including pure sliding,

shear-coupled migration, dislocation emission, and structural transformations that altered the

initial boundary structure. Reversing the initial loading direction frequently changed the shear

strength and, in some cases, the deformation mode. For instance, the {111}/{331}(49°) GB

exhibited shear-coupled migration in one direction and pure sliding in the reverse. These results

demonstrate a pronounced direction-dependent asymmetry in shear responses, shifting the active

11

deformation mechanism even within identical initial boundary configurations.

Fig. 3 | Atomic structures, structural periodicities, and shear responses of selected

asymmetric incommensurate GBs. a, Extended GB in a Ni bicrystal after energy minimization,

joining {331} and {111} terminating planes at a 22° misorientation angle, denoted as

{111}/{331}(22°). b, Magnified view of the interface. Small and large red regions near the {331}

terminating plane identify the cores of PGMDs and SGMDs, respectively. The primary period (𝜆𝜆 =

2.16 nm) is the spacing between the repeating arrangement of PGMDs, and the secondary period

(Λ = 15.73 nm) is the spacing between adjacent SGMDs. c, Magnified view of the boundary

showing adjacent PGMDs separated by one 𝜆𝜆. d, Periodic supercell constructed from the structure

in c and relaxed under PBCs to eliminate normal stress components for uniform sliding simulation.

Atoms in a-d are colored according to atomic volumetric strain. e, Athermal shear strengths of the

selected asymmetric incommensurate GBs loaded in opposite directions parallel to the GB plane

and perpendicular to the ⟨110⟩ tilt axis. Directions 1 and 2 denote the lower and higher shear

resistance, respectively. Symbol colors represent the observed deformation modes. f, Relaxed GB

12

energies per unit area for the compact boundary models in d.

To examine the relationship between GB energy and shear resistance, the LAMMPS agent

calculated the energies of all relaxed boundaries (Fig. 3f and Supplementary Note 1). Although

higher-energy boundaries generally exhibited a trend toward lower shear resistance, this

relationship was non-monotonic. For example, the {111}/{100} GB had both the lowest relaxed

GB energy and the lowest shear strength of 14 MPa. Furthermore, for most boundaries lacking

structural symmetry about the interface normal, loading in opposite directions yielded different

shear strengths and deformation modes despite identical GB energies. These findings demonstrate

that static GB energy alone is insufficient to predict or characterize direction-dependent shear

responses, underscoring the need to directly analyze atomic rearrangements and deformation

pathways under shear loading.

Pure sliding

Four GB configurations exhibited pure sliding under both shear directions: the {111}/{100},

{111}/{331}(22°), {100}/{110} and {111}/{110} GBs. To investigate their deformation pathways,

the analysis agent correlated stress histories and stress-drop events with corresponding atomic

rearrangements (Fig. 4). Each boundary contained at least one close-packed {111} or densely

packed {100} terminating plane, whereas differences in the opposing terminating planes produced

distinct interfacial registries. Specifically, the {110} termination exhibited atomic-scale

corrugation, whereas the {331} termination comprised single-atomic-layer steps separated by

short {111} facets. These geometric differences produced distinct local structural motifs within

each interfacial repeat unit. Table 2 lists the primary periodicities λ of all four boundaries.

Comparing atomic configurations across selected sliding intervals, particularly immediately

before and after stress drops, revealed the associated interfacial displacement fields (Fig. 4). Within

a single repeat unit, atomic displacements were either distributed across multiple atomic columns

or localized to individual atomic columns oriented along the ⟨110⟩ tilt axis. Atoms within adjacent

columns translated in different directions, indicating that macroscopically uniform tangential grain

translation was accommodated by locally nonuniform atomic shuffling. The spatial distribution of

these rearrangements varied markedly among the four pure-sliding interfaces. While both the

{111}/{100} and {111}/{110} GBs exhibited low shear strengths (Fig. 4a, b), their displacement

patterns differed: displacements were distributed more uniformly in the {111}/{100} GB than in

13

the {111}/{110} GB. Conversely, the {111}/{331}(22°) and {100}/{110} GBs exhibited

substantially higher shear strengths, yet also differed in their local displacement modes. In the

{111}/{331}(22°) GB, displacements were distributed across all atomic columns without a single

column dominating the response. In contrast, each stress-drop event in the {100}/{110} GB

involved pronounced shuffling localized within an individual column.

Fig. 4 | Supercell atomic structures, shear stress-strain responses, and displacement fields of

GBs exhibiting pure sliding. a1–a4, {111}/{100} GB; b1–b4, {111}/{110} GB; c1–c4,

{111}/{331}(22°) GB; d1–d4, {100}/{110} GB. For each boundary, panel 1 shows the relaxed

atomic structure in a compact supercell with PBCs applied in the boundary plane, with

14

crystallographic directions indicated in the two adjoining grains. Red arrows indicate the primary

interfacial repeat length λ. Panel 2 shows the shear stress-strain curve, where states (1)-(3) define a representative sliding interval. Panels 3 and 4 present GB plan views of the boundary at states

(2) and (3), respectively. Red arrows denote atomic displacement vectors from state (1) to (2) in

panel 3 and from state (2) to (3) in panel 4. Vector amplification factors are specified alongside

each configuration.

Table 2 | Structural and deformation characteristics of four pure-sliding GBs. Listed

quantities include the SGMD Burgers vector 𝐛𝐛 s , the primary periodicity vector 𝛌𝛌 and its magnitude |𝛌𝛌|, the average secondary periodicity |𝚲𝚲�|, and the GB shear strength without (𝜏𝜏 0) and with (𝜏𝜏 s) SGMDs. Both 𝐛𝐛 s and 𝛌𝛌 are given in units of the lattice constant 𝑎𝑎 0, with the slip plane expressed in terms of the terminated plane of the reference grain. Where two shear strengths are

listed, the upper and lower entries correspond to shear directions 1 and 2, respectively.

GB configuration

𝐛𝐛 s 𝛌𝛌

|𝛌𝛌| (nm)

|𝚲𝚲�|* (nm)

𝜏𝜏 0 (MPa)

𝜏𝜏 s (MPa)

{111}/{100}

1/14⟨112⟩{111} 2⟨112⟩{111} 1.73 5.97 14 3.5 × 10−4 1/8⟨110⟩{100} 7/2⟨110⟩{100}

{111}/{110}

1/22⟨112⟩{111} 9/2⟨112⟩{111} 3.87 19.0 23 1.4 × 10−2 1/9⟨100⟩{110} 11⟨100⟩{110}

{111}/{331} (22°)

1/8⟨112⟩{111} 5/2⟨112⟩{111} 2.16 16.25

800 -**

1/20⟨116⟩{331} ⟨116⟩{331} 1,074 530

{100}/{110}

1/24⟨110⟩{100} 17/2⟨110⟩{100} 4.23 11.96 3,498 3,870 1/17⟨100⟩{110} 12⟨100⟩{110}

* |𝚲𝚲�| represents the average spacing between neighboring SGMDs. Individual spacings are integer multiples of |𝛌𝛌| , typically 𝑛𝑛|𝛌𝛌| and (𝑛𝑛 + 1)|𝛌𝛌| , where 𝑛𝑛 is an integer 43. The reported value is the average of these spacings weighted by their relative frequencies. **The GB underwent successive structural transformations during shear loading, precluding determination of a well-defined shear strength.

Iterative workflow refinement enabled GBClaw to connect structural simulation data directly

with mechanistic analysis. Combining atomic displacement fields with structural recurrence

analysis linked stress drops to discrete grain translations, thereby identifying the displacement-

shift-complete (DSC) slip vector44 . For the {111}/{331}(22°) GB, each stress drop produced a

15

translation of 1/8⟨112⟩{111}, relative to the {111} terminated grain. Four successive drops yielded

a cumulative slip of 1/2⟨112⟩{111}, translating each atom on the {331} terminating plane to a

crystallographically equivalent site along the sliding direction. The grain translation associated

with a single stress drop represents the shortest displacement that preserves the interfacial structure,

thus defining the SGMD Burgers vector b s (Fig. 4c). Table 2 lists b s determined for all four GBs,

whose variations reflect their distinct bicrystallography. To validate these assignments, interactive

exchanges between the researcher and the orchestrator prompted the development of an automated

procedure to identify SGMDs from projected atomic-column arrangements in HRTEM images

(Supplementary Note 2, Figs. S4, S5 and Table S2). Subsequent exchanges further enabled an

independent crystallographic analysis using near-CSL dichromatic patterns to validate the

extracted Burgers vectors b s (Supplementary Note 3 and Fig. S6). Agreement among these

approaches corroborated the SGMD identification and Burgers vector assignments, establishing a

general geometrical procedure for analyzing SGMDs in experimental HRTEM images and

atomistic simulations. These iterative exchanges illustrate how researcher guidance enabled

GBClaw to refine its workflows and validate mechanistic interpretations through complementary

methods.

To examine how SGMDs affect GB sliding, the structure and LAMMPS agents evaluated the

shear strengths of bicrystal models containing arrays of these defects. Each supercell incorporated

multiple SGMDs whose Burgers vectors summed to a lattice repeat along the boundary, preserving

crystallographically equivalent interfacial registry across periodic images. Under PBCs, these

models represent extended GBs containing regular SGMD arrays. For the {111}/{331}(22°) GB,

sliding proceeded through the intermittent glide of individual dislocations or localized sub-arrays

(Supplementary Note 4 and Fig. S7). The presence of SGMDs substantially reduced the shear

strength from 𝜏𝜏 0 = 1,074 MPa in the SGMD-free compact model to 𝜏𝜏 s = 530 MPa in the model with SGMDs. Table 2 compares 𝜏𝜏 0 and 𝜏𝜏 s for all four GBs alongside the corresponding

magnitudes of the secondary periodicity vector Λ. These results demonstrate a significant disparity

in predicted shear strengths between GB models with and without SGMDs, underscoring the

necessity of incorporating defect hierarchies into atomistic models. A systematic analysis of how

spatially non-uniform distributions of SGMDs influence sliding in general GBs will be reported

separately.

Dynamic replanning enabled GBClaw to leverage researcher guidance to investigate an

anomalous shear response and resolve its structural origin. During the initial shear simulation of

16

the {111}/{110} GB, an unusually high initial stress was followed by a sharp drop and steady-state

sliding at a substantially lower stress. In response to a researcher’s inquiry, the orchestrator

dynamically expanded the simulation plan to execute a loading-unloading cycle and NEB

calculations (Fig. 5a and Supplementary Note 5). After unloading to zero applied shear strain, the

bicrystal retained a nonzero shear stress, whereas the initial configuration was stress free (inset in

Fig. 5a). The analysis agent then compared the two zero-strain configurations and mapped the

atomic displacements between them (Fig. 5b). This comparison showed that unloading did not

restore the initial interfacial structure. The two relaxed configurations instead occupied distinct

local minima under identical zero-strain constraints.

Fig. 5 | Loading-induced local reconstruction and associated energy barrier in the {111}/{110}

GB. a. Shear stress-strain response during loading (blue) and unloading (red). The inset shows the

residual shear stress of the initial relaxed configuration (1) and the final unloaded configuration

(2), both evaluated at zero applied shear strain. b. Edge-on (upper) and plan (lower) views of

configuration (2). Red arrows indicate atomic displacement vectors from configuration (1) to (2),

scaled by a factor of two. c. Minimum energy path from an NEB calculation connecting

configurations (1) and (2) at zero applied shear strain. d. Atomic configuration at the saddle point

in c. Atoms are colored by their displacement magnitude relative to configuration (1).

17

To characterize the energy landscape connecting these states, GBClaw performed an NEB

calculation using the relaxed initial and reconstructed GB configurations as endpoints at zero

imposed shear strain. The resulting minimum energy path exhibited an energy barrier of

approximately 20 meV/𝑎𝑎 0, where 𝑎𝑎 0 = 3.52 Å is the lattice constant, and terminated at a state with lower energy than the initial configuration (Fig. 5c). This result indicated the metastability of

the initial GB structure. Together with the localized atomic displacements at the saddle-point state

relative to the initial state (Fig. 5d), these results revealed a loading-induced local reconstruction

of the {110} terminating plane against the opposing {111} plane. These findings attributed the

initially high sliding resistance to a metastable local structure formed during bicrystal construction

and relaxation. Reloading the reconstructed zero-strain configuration did not reproduce the

anomaly and instead yielded the normal stress-strain response shown in Fig. 4b2. This

investigation demonstrates how human-in-the-loop replanning empowers GBClaw to unravel

unexpected simulation phenomena and pinpoint their atomistic reconstruction mechanisms.

Shear asymmetry between GB sliding and migration

The analysis agent compared the 22° and 49° misorientation variants of the {111}/{331} GB,

providing mechanistic insights into shear asymmetry and deformation-mode selection. The two

variants arise from two {111} planes whose inclination angles differ by 71° about the ⟨110⟩ tilt

axis. Both variants share a primary periodicity λ with a magnitude of 2.16 nm and an SGMD

Burgers vector bs = 1/8⟨112⟩{111} (Table 2, Table S4 and Supplementary Fig. S8). When sheared

along direction 1 (the mode of lower shear strength), the 49° variant exhibited a pure-sliding

response similar to that of the 22° variant. Its repeated stress drops produced a comparable

sawtooth profile, albeit at different stress levels (Figs. 4c2 and 6a). Atomic configurations and

displacement vectors likewise confirmed tangential grain translation without GB migration (Figs.

4c and 6c1-c3). When shear was applied along direction 2, however, the 49° GB underwent shear-

coupled migration, whereas the 22° GB maintained pure sliding. Specifically, stress-drop events

in the 49° GB coincided with downward boundary motion through the transformation of atomic

columns into the opposing grain orientation and coordinated shuffling across two atomic layers

(Fig. 6b, d1-d3).

Dichromatic mapping based on near-CSL constructions revealed different lattice registries for

the two variants (Supplementary Figs. S6c and S8a), suggesting that differences in subsurface

stacking sequences and second-nearest-neighbor coordination likely underlie the distinct shear

18

pathways. Supplementary Note 3 examines the shear asymmetry, while Supplementary Notes 6

and 7 detail the corresponding shear-coupled migration pathways and crystallographic mapping,

respectively. This comparison further illustrates how GBClaw bridges crystallographic

characterization with the local atomic mechanisms governing shear responses.

Fig. 6 | Shear asymmetry in the {111}/{331}(49°) GB. a, b, Shear stress as a function of applied

shear strain along directions 1 and 2, respectively. Insets show the initial boundary structures and

shear directions. c1-c3, Pure GB sliding when sheared along direction 1. d1-d3, Shear-coupled GB

migration when sheared along direction 2. c1, d1 and c2, d2 show configurations before and after,

respectively, the stress drops indicated in a and b. c3, d3, Corresponding atomic displacement

vectors, amplified twofold for clarity. Selected atomic rows are colored to highlight structural

changes. Black lines mark terminating planes. The magenta arrow indicates downward boundary

19

migration during shear along direction 2. All configurations are viewed along the ⟨110⟩ tilt axis.

Discussion

GBClaw combines automated workflows, LLMs, and domain expertise to bridge systematic

GB characterization with targeted mechanistic investigation. By coordinating structural screening,

shear simulations, and atomistic analyses, the framework identified misfit-dislocation hierarchies,

distinguished diverse deformation modes, and quantified substantial variations in athermal shear

strengths across asymmetric incommensurate GBs. Researcher-guided iterative workflows

enabled complementary analyses of SGMD structures and Burgers vectors, establishing general

procedures to identify and analyze SGMDs in both experimental images and atomistic simulations.

Atomistic calculations revealed a significant disparity in predicted shear strengths between GB

models with and without SGMDs. These results underscore the necessity of incorporating defect

hierarchies when applying high-fidelity, computationally intensive first-principles or machine-

learned potentials to design the composition and structure of high-performance general GBs.

Furthermore, dynamic replanning directed follow-up calculations that resolved the structural

origins of the anomalous shear response of the {111}/{110} GB and shear asymmetry of the

{111}/{331}(49°) GB. These examples demonstrate how automation scales systematic exploration

while domain expertise guides workflow development, validation, and physical interpretation.

Several methodological constraints of the underlying atomistic simulations warrant

consideration. The current implementation employed compact supercells with PBCs within the

GB plane to evaluate shear strengths with and without SGMDs. These spatial constraints restrict

the spatial distribution and long-range interactions of line defects, potentially suppressing

deformation pathways active in extended boundaries. The calculated strengths therefore represent

athermal instability thresholds for the modeled boundary configurations, rather than universal

upper bounds. Likewise, while NEB calculations delineate minimum energy paths between

selected states, they do not explicitly account for finite-temperature dynamics. Addressing these

limitations will require large-scale molecular dynamics simulations to capture interactions among

multiple defects, alongside enhanced sampling or kinetic rate-theory methods to evaluate

thermally activated processes.

GBClaw’s modular architecture and hierarchical memory provide an extensible foundation

for addressing these challenges. Specialized agents currently under development focus on the

adaptive refinement of machine-learned potentials tailored to complex GB configurations and on

20

experimental validation through atomic-resolution characterization. Integrating these agents

through standardized data interfaces will enable GBClaw to coordinate high-throughput structural

screening, potential refinement, atomistic simulation, and experimental validation within a single

closed-loop system.

To balance flexibility and reliability, GBClaw combines standardized scientific tools with

custom code dynamically generated by its coding agent. Predefined tools establish reproducible

workflows for routine operations 45 , whereas custom code provides the adaptability needed for

specialized analyses and error recovery 16 . However, token consumption aside, repeated code

generation risks introducing algorithmic inconsistencies across runs 46 . Converting validated scripts

into persistent, structured tools after their physical assumptions, inputs, and numerical behavior

have been checked mitigates this variability. As highlighted by the domain-knowledge pitfalls

reported for MatClaw6 , successful execution does not guarantee physical correctness. Multi-layer

verification, physical sanity checks, and human expert oversight remain indispensable.

In summary, GBClaw demonstrates that agentic AI can effectively connect high-throughput

simulations to deep mechanistic discovery, while maintaining full provenance across models,

execution records, and analytical outputs. By elucidating the governing role of misfit-dislocation

hierarchies, this work advances a unified, defect-based framework for investigating complex

interfaces. Within general asymmetric tilt GBs, PGMDs are typically sessile in the boundary plane,

acting as structural obstacles to motion of other defects. In contrast, SGMDs serve as mobile

carriers of interfacial sliding, forming either to accommodate interfacial lattice mismatch or

through the transformation of dislocations absorbed from the grain interior or triple junctions.

Beyond homophase boundaries, GBClaw provides a generalizable framework for investigating

heterophase interfaces, van der Waals heterostructures, and interfacial friction in incommensurate

solid-state contacts.

Methods

Implementation of GBClaw

GBClaw was implemented in Python using the open-source LangChain and LangGraph

frameworks 17 for agent coordination and DeepAgents for task-specific code execution. The system

used both GPT-5.6 Sol and GPT-5.6 Luna, with task-dependent model routing selecting the model

according to the complexity and reasoning requirements of each task. Specialized agents accessed

LAMMPS 25 , OVITO26 , and ASE27 through Python-based tool interfaces. Computational results

21

were linked to persistent research records and supporting artifacts. Supplementary Table S3 lists

the tools available to each agent, and Supplementary Fig. S3 illustrates a representative researcher-

guided investigation. Further implementation details and software dependencies are documented

in the source code repository identified in the Code Availability statement.

GB selection and model construction

General asymmetric ⟨110⟩ tilt GBs were selected near the exact CSL reference boundaries

with Σ < 50 listed in Table 1. Crystallographically equivalent plane pairs were grouped according

to cubic symmetry, and one representative orientation was retained from each group. The

{111}/{331}(49°) and {111}/{221}(55°) variants were added to the 12 selected plane-family pairs,

yielding 14 GB configurations. Ni bicrystals were constructed with the tilt axis along the out-of-

plane direction. Interatomic interactions were described using the angular-dependent Ni-Cr

potential of Howells and Mishin41 . The lattice parameter of Ni a0 was 3.52 Å. Compact periodic

supercells were selected with an integer multiple of the primary periodicity along the horizontal

direction, as illustrated in Fig. 3. These supercells retained the local interfacial motif while

excluding SGMDs. Residual mismatch was accommodated by elastic strain, with both the atomic

coordinates and cell lengths allowed to relax. Supplementary Table S4 reports the crystallographic

orientations, lattice repeat counts, relaxed supercell dimensions and residual lattice mismatch.

Each compact supercell contained one GB between upper and lower free surfaces. PBCs were

applied within the GB plane, and vacuum regions with an initial thickness of 20 Å were introduced

above and below the free surfaces. Before relaxation, overlapping atoms separated by less than

0.3a0 across the GB were removed. Structural relaxation was performed in LAMMPS25 through

four cycles, each alternating between atomic relaxation at fixed supercell dimensions and cell-

length relaxation toward zero normal stress. Conjugate gradient energy minimization used a force

tolerance of 10-6 eVÅ -1 , with maximum limits of 1,000 iterations and 10,000 force evaluations for

each minimization. The energy-based convergence criterion was disabled.

Shear simulations

Shear displacement was applied incrementally along the horizontal direction. Boundary layers

with a thickness of 3 Å were defined at the upper and lower surfaces of the bicrystal, and their

atomic membership was retained throughout loading. At each load increment, the upper layer was

displaced horizontally while the lower layer remained fixed in that direction. During energy

minimization, the in-plane coordinates of both boundary layers were constrained, while their

22

coordinates normal to the GB plane were allowed to relax. The cell dimensions remained fixed

throughout loading. The shear strain was defined as γ = Δu/H0 , where Δu is the relative horizontal

displacement between the centers of the boundary layers, and H0 is their initial separation normal

to the GB plane.

The default forward displacement used a strain increment of Δγ = 2.5×10-4 over 200

increments. Each increment was followed by conjugate gradient energy minimization with a force

tolerance of 10-6 eVÅ-1 . The energy-based convergence criterion was disabled, and each

minimization was limited to 3,000 iterations and 10,000 force evaluations. Shear stress was

calculated as the negative of the LAMMPS pressure-tensor component Pxy and rescaled by the

ratio of the initial supercell volume to the initial atomic bounding-box volume, thereby excluding

the vacuum contribution. The shear strength in Fig. 3e was defined by the peak stress within the

subsequent steady response after excluding the initial loading transient. Directions 1 and 2 denote

the directions with lower and higher shear resistance, respectively. For the {111}/{100} and

{111}/{110} GBs shown in Fig. 4a, b, the strain increment was reduced fivefold to 5×10 -5 to

resolve the sliding response more finely. Simulations with periodic arrays of SGMDs are described

in Supplementary Note 4 and Supplementary Fig. S7.

Loading-unloading cycle and NEB calculations

The {111}/{110} GB was loaded beyond the initial stress drop and unloaded to zero applied

shear strain using the same displacement-control procedure. The loading range and number of

increments are provided in Supplementary Note 5. The relaxed initial and unloaded configurations

were used as endpoints for a strain-controlled NEB calculation 22,23 at zero applied shear strain.

Atoms were matched by their identifiers, and 144 replicas, including both endpoints, were

initialized through linear interpolation using periodic minimum-image displacements. The

minimum energy path was obtained using the FIRE algorithm with a spring coefficient of 1.0 eVÅ-

2 and a minimization timestep of 0.01 ps. The conventional NEB and climbing-image stages were

each limited to 10,000 iterations and used a force tolerance of 10 -6 eVÅ-1 . The energy-based

convergence criterion was disabled. Energies were reported relative to the relaxed initial endpoint.

Supplementary Table S4 provides supercell dimensions.

Code availability

The GBClaw source code is provided under the MIT License on GitHub

23

(https://github.com/ZhuGroup-GT/GBClaw). The recorded demo project can be visited at

https://zhugroup.gatech.edu/gbclaw-demo.

Data availability

All relevant data are included within the manuscript, Supplementary Information, and GitHub repository. The detailed simulation artifacts are available from the corresponding author upon request.

Author contributions

YL and TZ conceived research. YL, KD, XG, SO and TZ conducted research. YL wrote the initial manuscript. WC and TZ revised the manuscript. All the authors contributed to the discussion of the results.

Competing interests

The authors declare no competing interests.

Acknowledgements

YL, KD, YY, JK, OP and TZ acknowledge support by the U.S. Department of Energy (DOE),

Office of Science, Basic Energy Sciences (BES) Materials Science and Engineering (MSE)

Division under Award #DE-SC0018960. WC acknowledges support by the National Science

Foundation (Award numbers 2516550 and 2331132). XG and SO acknowledge support by Japan

Society for the Promotion of Science (JSPS) KAKENHI Grant No. JP23K20037. SO

acknowledges support by Japan Science and Technology Agency (JST) ARiSE Grant Numbers

JPMJAR2601 and JPMJAR2603 and Ministry of Education, Culture, Sport, Science and

Technology of Japan (MEXT) Program Grant Number JPMXP1122684766.

24

References

1. Cheng, M. et al. Artificial intelligence-driven approaches for materials design and discovery. Nat. Mater. 25, 174–190 (2026). 2. Ghareeb, A. E. et al. A multi-agent system for automating scientific discovery. Nature 655, 497–505 (2026). 3. Wei, J. et al. From AI for Science to Agentic Science: A Survey on Autonomous Scientific Discovery. Preprint at https://doi.org/10.48550/arXiv.2508.14111 (2025). 4. Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023). 5. Ifargan, T., Hafner, L., Kern, M., Alcalay, O. & Kishony, R. Autonomous LLM-Driven Research - from Data to Human-Verifiable Research Papers. NEJM AI 2, AIoa2400555 (2025). 6. Zhang, C. & Yakobson, B. I. MatClaw: An Autonomous Code-First LLM Agent for End-to- End Materials Exploration. Preprint at https://doi.org/10.48550/arXiv.2604.02688 (2026). 7. Gottweis, J. et al. Accelerating scientific discovery with Co-Scientist. Nature 655, 487–496 (2026). 8. Coble, R. L. A Model for Boundary Diffusion Controlled Creep in Polycrystalline Materials. J. Appl. Phys. 34, 1679–1682 (1963). 9. Watanabe, T. & Tsurekawa, S. The control of brittleness and development of desirable mechanical properties in polycrystalline systems by grain boundary engineering. Acta Materialia 47, 4171–4185 (1999). 10. Lu, K., Lu, L. & Suresh, S. Strengthening Materials by Engineering Coherent Internal Boundaries at the Nanoscale. Science 324, 349–352 (2009). 11. Sutton, A. P. & Balluffi, R. W. Interfaces in Crystalline Materials. (Clarendon Press, 1995). 12. Seki, T. et al. Incommensurate grain-boundary atomic structure. Nat Commun 14, 7806 (2023). 13. Woguem, Y. F., Godard, P., Durinck, J. & Brochard, S. Atomic scale simulations of {112} symmetric incoherent twin boundaries in gold. Materialia 27, 101678 (2023). 14. Cahn, J. W., Mishin, Y. & Suzuki, A. Coupling grain boundary motion to shear deformation. Acta Materialia 54, 4953–4975 (2006). 15. Wan, L. & Li, J. Shear responses of [\bar1\,1\,0]-tilt 1 1 5/1 1 1 asymmetric tilt grain boundaries in fcc metals by atomistic simulations. Modelling Simul. Mater. Sci. Eng. 21, 055013 (2013). 16. Wang, X. et al. Executable Code Actions Elicit Better LLM Agents. in Proceedings of the 41st International Conference on Machine Learning vol. 235 50208–50232 (PMLR, 2024). 17. Chase, H. LangChain. https://github.com/langchain-ai/langchain (2022). 18. Grimmer, H., Bollmann, W. & Warrington, D. H. Coincidence-site lattices and complete pattern-shift in cubic crystals. Acta Cryst A 30, 197–207 (1974). 19. Gautam, A., Ophus, C., Lançon, F., Radmilovic, V. & Dahmen, U. Atomic structure characterization of an incommensurate grain boundary. Acta Materialia 61, 5078–5086 (2013). 20. Hadian, R., Grabowski, B. & Neugebauer, J. GB code: A grain boundary generation code. Journal of Open Source Software 3, 900 (2018). 21. Ding, K. et al. Misfit-dislocation hierarchy governs sliding of asymmetric non-CSL grain

25

boundaries. Preprint at https://doi.org/10.48550/arXiv.2609.14673 (2026). 22. Jonsson, H., Mills, G. & Jacobsen, K. W. Nudged elastic band method for finding minimum energy paths of transitions. in Classical and Quantum Dynamics in Condensed Phase Simulations 385–404 (World Scientific, 1998). doi:10.1142/9789812839664_0016. 23. Zhu, T., Li, J. & Yip, S. Atomistic Reaction Pathway Sampling: The Nudged Elastic Band Method and Nanomechanics Applications. in Nano and Cell Mechanics 311–338 (John Wiley & Sons, Ltd, 2013). doi:10.1002/9781118482568.ch12. 24. Jin, Q. et al. AgentMD: Empowering language agents for risk prediction with large-scale clinical tool learning. Nat Commun 16, 9377 (2025). 25. Thompson, A. P. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications 271, 108171 (2022). 26. Stukowski, A. Visualization and analysis of atomistic simulation data with OVITO–the Open Visualization Tool. Modelling Simul. Mater. Sci. Eng. 18, 015012 (2010). 27. Hjorth Larsen, A. et al. The atomic simulation environment-a Python library for working with atoms. J. Phys.: Condens. Matter 29, 273002 (2017). 28. Sumers, T. R., Yao, S., Narasimhan, K. & Griffiths, T. L. Cognitive Architectures for Language Agents. Preprint at https://doi.org/10.48550/arXiv.2309.02427 (2024). 29. Liu, N. F. et al. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics 12, 157–173 (2024). 30. Saylor, D. M., El Dasher, B. S., Rollett, A. D. & Rohrer, G. S. Distribution of grain boundaries in aluminum as a function of five macroscopic parameters. Acta Materialia 52, 3649–3655 (2004). 31. Randle, V., Rohrer, G. S., Miller, H. M., Coleman, M. & Owen, G. T. Five-parameter grain boundary distribution of commercially grain boundary engineered nickel and copper. Acta Materialia 56, 2363–2373 (2008). 32. Beladi, H. & Rohrer, G. S. The Distribution of Grain Boundary Planes in Interstitial Free Steel. Metall Mater Trans A 44, 115–124 (2013). 33. Hardouin Duparc, O. et al. Atomic structures of symmetrical and asymmetrical facets in a near Σ=9{221} tilt grain boundary in copper. Acta Materialia 55, 1791–1800 (2007). 34. Bowers, M. L., Ophus, C., Gautam, A., Lançon, F. & Dahmen, U. Step Coalescence by Collective Motion at an Incommensurate Grain Boundary. Phys. Rev. Lett. 116, 106102 (2016). 35. Wang, L. et al. Tracking the sliding of grain boundaries at the atomic scale. Science 375, 1261– 1265 (2022). 36. Tian, Y. et al. Grain rotation mechanisms in nanocrystalline materials: Multiscale observations in Pt thin films. Science 386, 49–54 (2024). 37. Medlin, D. L., Cohen, D. & Pond, R. C. Accommodation of coherency strain by interfacial disconnections at a 90° ⟨110⟩ grain boundary in gold. Phil Mag Lett 83, 223–232 (2003). 38. Han, J., Thomas, S. L. & Srolovitz, D. J. Grain-boundary kinetics: A unified approach. Progress in Materials Science 98, 386–476 (2018). 39. Brandon, D. G. The structure of high-angle grain boundaries. Acta Metallurgica 14, 1479–

26

1484 (1966). 40. Brokman, A. & Balluffi, R. W. Coincidence lattice model for the structure and energy of grain boundaries. Acta Metallurgica 29, 1703–1719 (1981). 41. Howells, C. A. & Mishin, Y. Angular-dependent interatomic potential for the binary Ni-Cr system. Modelling Simul. Mater. Sci. Eng. 26, 085008 (2018). 42. Lançon, F. et al. Superglide at an Internal Incommensurate Boundary. Nano Lett. 10, 695–700 (2010). 43. Friedel, J. & de Gennes, P.-G. Friction between incommensurate crystals. Philosophical Magazine 87, 39–49 (2007). 44. Cai, W. & Nix, W. D. Imperfections in Crystalline Solids. (Cambridge University Press, Cambridge, 2016). doi:10.1017/CBO9781316389508. 45. Huber, S. P. et al. Common workflows for computing material properties using different quantum engines. npj Comput Mater 7, 136 (2021). 46. Ouyang, S., Zhang, J. M., Harman, M. & Wang, M. An Empirical Study of the Non- Determinism of ChatGPT in Code Generation. ACM T Softw Eng Meth 34, Article 42 (2025).

27

Supplementary Information

GBClaw: An agentic AI framework for atomistic grain boundary research

and mechanistic discovery

Yazhuo Liua,† , Kunqing Dinga,† , Xiaoguo Gongb , Zifeng Wang a , Yichen Yang a , Yifeng Chea , Wen

Chend,e , Josh Kacherc , Olivier Pierron a , Shigenobu Ogatab , Ting Zhua,c, *

a Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA,

30332, USA

b Department of Mechanical Science and Bioengineering, Graduate School of Engineering

Science, The University of Osaka, Osaka, 560-8531, Japan

c School of Materials Science and Engineering, Georgia Institute of Technology, Atlanta, GA,

30332, USA

d Department of Aerospace and Mechanical Engineering, University of Southern California, Los

Angeles, CA, USA

e Mork Family Department of Chemical Engineering and Materials Science, University of

Southern California, Los Angeles, CA, USA

† These authors contributed equally to this work.

* Corresponding author: ting.zhu@me.gatech.edu

28

Supplementary figures

Fig. S1 | Memory organization and identifier-based information flow in GBClaw.

Conversation state retains the session summary, recent messages, and recent tool inputs and

outputs. Semantic memory stores project plans, conclusions, facts, run records and an artifact index.

Shared run IDs connect fact records with simulation conditions and outputs, while artifact IDs link

indexed entries to the corresponding files for retrieval on demand. Complete interaction and tool

execution histories are archived for human review. Green and blue arrows represent summarized

and raw information, respectively. Orange arrows denote memory links, and magenta arrows

indicate artifact queries.

29

Fig. S2 | Detailed GBClaw workflow for investigating GB shear strength. Automated tasks and

researcher guidance connect crystallographic screening, simulation planning, parallel bicrystal

construction and relaxation, selection of periodic boundary conditions, shear simulations, and

structural and mechanical analysis of general asymmetric ⟨110⟩ tilt GBs. The {110}/{111} GB

illustrates an adaptive investigation, in which an unexpected shear response prompted researcher-

directed follow-up analyses using cyclic shear simulations and NEB calculations.

30

Fig. S3 | Representative researcher-GBClaw interaction underlying the workflow in Fig. 2.

a, Selection of bicrystal orientations and periodic-cell candidates. b, Construction, relaxation and

visualization of the selected bicrystal. c, Unloading to zero applied strain and comparison with the

initial state. d, NEB calculation between the initial and unloaded configurations. e, Near-CSL

dichromatic analysis of candidate SGMD Burgers vectors. f, Relaxation of selected configurations

and comparison of their displacement fields. g, Atomic displacements during a selected unloading

interval. Detailed tool calls, reasoning summaries and computational outputs are available in the

accompanying online demonstration (https://zhugroup.gatech.edu/gbclaw-demo). The displayed

NEB setup is illustrative. Final NEB settings are given in Methods and Supplementary Note 5.

31

Fig. S4 | Identification of SGMDs from projected atomic-column arrangements in simulated

structures. a, {111}/{100} GB. b, {111}/{110} GB. c, {111}/{331}(22°) GB. d, {100}/{110} GB.

All configurations are viewed along the common ⟨110⟩ tilt axis. White lines trace edge-on atomic

planes, and yellow markers identify their interfacial terminations. Red double-headed arrows

indicate the primary structural periodicity λ. Green double-headed arrows span the altered row

correspondence within SGMD cores. Comparing core and reference row counts gives the Burgers

vector magnitudes listed in Table S2, consistent with Table 2 and the near-CSL analysis in Fig. S6.

32

Fig. S5 | Identification of SGMDs from projected atomic-column arrangements in in situ HRTEM images of a {111}/{331}(22°) GB undergoing sliding, adapted from Wang et al 35 . a, b,

Each image shows two PGMDs. In b, solid white lines trace extra {111} half-planes, whose

interfacial terminations identify the PGMD cores. c, An SGMD identified by an additional {111}

half-plane. The associated atomic-column arrangement is consistent with the simulated structure

shown in Fig. S4c.

33

Fig. S6 | Determination of SGMD character from near-CSL dichromatic patterns of the four

pure-sliding GBs. a, {111}/{100} GB; b, {111}/{110} GB; c, {111}/{331}(22°) GB; d,

{100}/{110} GB. Blue and orange symbols distinguish the two grain lattices, while circles and

diamonds distinguish alternating atomic layers along the ⟨110⟩ tilt axis. Black dashed lines indicate

the near-CSL cells and GBs. Black arrows indicate the shortest DSC translations connecting lattice

sites in the two grains within the same atomic layer. These translations define the SGMD Burgers

vectors and agree with those obtained from the projected atomic-column analysis in Fig. S4 and

Table S2. Red dashed circles mark symmetry-related site pairs that produce DSC translations equal

in magnitude but opposite in sign to those indicated in red solid boxes.

34

Fig. S7 | Boundary sliding through nonuniform glide of SGMDs in the {111}/{331}(22°) GB.

a, Shear stress as a function of applied shear strain. b, Magnified view of the stress response, with

states (1) to (5) marking a representative sliding cycle. c1 to c5, Atomic configurations

corresponding to the marked states. Large red regions identify SGMD cores. Yellow lines mark

equivalent SGMD positions at the beginning and end of the cycle. Green lines track dislocation

positions, and arrows indicate movement between consecutive configurations.

35

Fig. S8 | SGMD Burgers vector derived from the near-CSL pattern of the {111}/{331}(49°)

GB. a, Near-CSL dichromatic pattern. Blue and orange symbols distinguish between the two grain

lattices, while circles and diamonds distinguish atomic layers along the ⟨110⟩ tilt axis. Dashed

outlines indicate the near-CSL cell. The black arrow in the inset marks the shortest DSC vector

connecting sites of the two lattices, yielding b s = 1/8⟨112⟩{111}, identical to that of the 22° boundary in Fig. S6c. The dashed circle marks a symmetry-related pair of sites with opposite

relative offsets in the reference pattern. b, Local interfacial configurations, with the blue and

orange grains positioned above and below the GB, respectively. Taking the middle configuration

as the reference, the orange atoms are held fixed, while the blue atoms are translated left by -bs to obtain the upper configuration and right by +bs to obtain the lower configuration. Green and red arrows denote these opposite tangential translations, each equal in magnitude to the Burgers vector

of the SGMD (bs). Black arrows indicate the resistance from second-nearest-neighbor atoms beneath the terminating plane. Yellow triangles highlight coincidence lattice sites, and vertical

dashed guides provide positional references across the three registries.

36

Fig. S9 | Spatially periodic atomic shuffling during shear-coupled GB migration. Shear stress

as a function of applied shear strain for a, {111}/{331}(49°) GB loaded in direction 2 and b,

{100}/{223} GB. Insets show the corresponding initial GB structures. Atomic configurations of

the {111}/{331}(49°) GB, c1, before and, c2, after the selected migration interval. d1-d2,

Corresponding atomic configurations of the {100}/{223} GB. Selected atomic sites are shown in

distinct colors to facilitate structural tracking. Black lines trace the terminating atomic planes,

outlined crystallographic cells indicate the orientations of the two grains, and magenta arrows mark

the directions of GB migration. c3, d3, Differential atomic displacement vectors from c1 to c2 and

from d1 to d2, respectively. The red vectors are amplified by factors of 2 and 0.75, respectively.

All atomic configurations are viewed along the ⟨110⟩ tilt axis.

37

Fig. S10 | Analysis of near-CSL dichromatic patterns associated with shear-coupled

migration. a, {111}/{331}(49°) GB. b, {100}/{223} GB. Colors distinguish the grain lattices, and

symbol shapes distinguish layers along the ⟨110⟩ tilt axis. Dashed lines indicate the near-CSL cells

and GB traces. Red arrows mark DSC vectors selected by comparison with the signed tangential-

to-normal displacement ratios measured during migration (Fig. S9). Crystallographic directions

are referenced to the grains indicated in each panel.

38

Table S1 | Hierarchical memory layers and access policies in GBClaw.

Level Layer Category Purpose Storage Agent access Update

0 Domain knowledge Procedural memory

Encode domain-specific workflows, tool contracts, and task execution rules

Built into tools and instructions

Implicitly accessed through tool invocation

Manual update only

1 Project summary Working memory

Give the orchestrator bounded project intent and conclusions for planning the next step

Working context and database file

Always in orchestrator context

Incrementally updated

2 Conversation state Episodic memory

Keep track of recent communication, decisions, and unresolved tasks

Working context and database file

Always in orchestrator context

Periodically compressed and refreshed

3 Facts/Runs/Conclusions Semantic memory

Store durable facts, conclusions, run summaries, and pointers to artifacts

Database file

On-demand retrieval

Updated with provenance

4 Artifacts External grounding

Store raw outputs, figures, trajectories, logs, code, and other files for traceable evidence

File system in artifact directory

On-demand retrieval

Persistent unless manually deleted by user

5 Full Chat/work history Human reference

Preserve the full append-only record for audit and human review

Database file

Excluded from agent retrieval Append-only

39

Table S2 | Determination of Burgers vectors 𝐛𝐛 𝐬𝐬 of SGMDs in Fig. S4.

GB pairs 𝑁𝑁u 𝑁𝑁l 𝑛𝑛 u 𝑛𝑛 l 𝛿𝛿 Δ𝑁𝑁 𝐝𝐝 𝐥𝐥/𝑎𝑎0* 𝐛𝐛𝐬𝐬/𝑎𝑎0*

{111}/{100} 7 8 5 6

2 35

2 7

1 4

⟨112⟩{111}

1 14

⟨112⟩{111}

{111}/{110} 11 18 5 8 −

2 55

−

2 11

1 4

⟨112⟩{111}

1 22

⟨112⟩{111}

{111}/{331}(22°) 8 10 2 3

1 4

1 2

1 4

⟨112⟩{111}

1 8

⟨112⟩{111}

{100}/{110} 12 17 5 7 −

1 60

−

1 12

1 2

⟨110⟩{100}

1 24

⟨110⟩{100}

* 𝑎𝑎 0 denotes the lattice constant. Vector coefficients are reported without signs, and the terminating plane specifies the reference grain. Here 𝛿𝛿 is the difference between the lower-to-

upper interval ratios in the core and reference spans, and Δ𝑁𝑁 is the accumulated mismatch defined

in Note 2. Positive and negative Δ𝑁𝑁 indicate an excess and a deficit of lower-grain row intervals

relative to the reference ratio, respectively.

40

Table S3 | Tools available to the orchestrator and specialized agents in GBClaw.

Agents Tool name(s) Usage

Orchestrator

list_available_potentials

Identify interatomic interaction models available for

simulations.

list_coding_environments

Determine which execution environments are available

for custom code.

search_web

Collect online scientific information relevant to the

investigation.

delegate_*(subagents) Route research tasks to the appropriate specialist agents.

get_project_plan

get_project_plan_draft

Recover the current research workflow and its pending

draft.

approve_project_plan_draft

reject_project_plan_draft

Accept or reject a proposed workflow before

implementation.

update_project_plan_step Record the status of individual workflow steps.

report_artifact_usage

Report the space occupied by stored computational

artifacts.

Planning

create_project_plan_draft

Translate research objectives into an ordered,

reviewable task plan.

get_project_plan

revise_project_plan_draft

Recover the current plan and revise proposed research

steps.

Structure

analyze_tilt_geometry

Establish the common tilt axis and relative grain

rotation.

compare_miller_planes

Assess the relative orientations of specified

crystallographic planes.

match_csl_tilt_references

Associate a boundary geometry with nearby exact CSL

misorientations.

score_csl_tilt_plane_pairs

Assess candidate CSL terminating-plane pairs for

selection.

suggest_low_index_near_csl_planes

Find low-index terminations close to a reference CSL

geometry.

list_csl_sigmas

List available CSL misorientations and their rotation

branches.

41

Agents Tool name(s) Usage

list_csl_gb_planes

list_csl_gb_planes_batch

Obtain terminating-plane pairs for individual or multiple

CSL references.

get_tilt_gb_orientation

Define grain coordinate systems and compatible lattice

repeat lengths.

get_dichromatic_map_options

render_dichromatic_map

Set visualization options and display superposed grain

lattices.

prepare_dichromatic_cell_selection

get_dichromatic_cell_selection

Prepare and recover a user-selected dichromatic unit

cell.

apply_dichromatic_cell_strain

Determine the strain required to match a selected near-

CSL cell.

analyze_dichromatic_cell_edges

Determine displacement vectors associated with

selected cell edges.

query_csl_orientation Recover previously stored CSL crystallographic data.

LAMMPS

list_available_potentials

Identify interatomic interaction models available for

simulations.

convert_cfg_to_lammps

Prepare LAMMPS input structures from CFG atomic

configurations.

infer_tilt_orientation

Establish grain orientations from specified boundary-

plane pairs.

estimate_gb_box_size

Estimate simulation-cell dimensions for bicrystal

construction.

write_gb_minimize_config

run_gb_minimize

Set construction parameters and obtain relaxed bicrystal

configurations.

write_structure_minimize_config

run_structure_minimize

Set relaxation parameters and minimize an existing

atomic configuration.

write_gb_energy_config

run_gb_energy

Set calculation parameters and evaluate interfacial

excess energy.

write_gb_shear_config

run_gb_shear

Define loading conditions and calculate the boundary

shear response.

write_gb_neb_config

run_gb_neb

Define transition endpoints and determine minimum

energy paths.

42

Agents Tool name(s) Usage

Analysis

analyze_gb_periodicity

Measure structural repeat lengths along the grain

boundary.

stress_strain_analysis

Locate stress drops and associated deformation events

in shear histories.

plot_stress_strain_comparison Display shear responses together for direct comparison.

analyze_neb_mep

plot_neb_mep

Extract transition barriers and visualize the calculated

energy path.

render_atomic_structure Produce atomic-structure images through OVITO.

render_gb_plan_view

Display the interface along its normal for structural

inspection.

Coding

ls; glob; grep

Find relevant files and locate information within their

contents.

read_file; write_file; edit_file; delete

Inspect, create, modify or remove files within the coding

sandbox.

execute Run custom commands within the coding sandbox.

publish_coding_result

Validate coding outputs and make them available to the

workflow.

All agents

query_project_memory

search_project_memory

Find project records relevant to the current task.

read_agent_conclusion

get_task_state

read_tool_output

Recover agent findings, task progress and complete tool

results.

All except

Coding

list_run_artifacts

read_run_artifact

Find output files associated with a run and inspect their

contents.

43

Table S4 | Grain orientations and dimensions of compact periodic supercells for 14 selected

general asymmetric GBs. The x-direction follows the applied shear direction, the y-direction is

normal to the GB, and the z-direction follows the common ⟨110⟩ tilt axis. Primary lattice repeats

𝛌𝛌 , cell sizes and residual lattice mismatches are listed. Crystallographic repeat vectors are

expressed in units of 𝒂𝒂𝟎𝟎 and referenced to the corresponding grain orientations.

GB plane pairs

Orientation of grain 1

(x, y, z)

Orientation of grain 2

(x, y, z)

𝛌𝛌

(Grain 1, grain 2)

Supercell size*

(nm3 )

Lattice

mismatch

{100}/{110} [001�], [1�10], [110] [11�0], [001�], [110] 12[001�],

17 2

[11�0] 8.45×23.95×1.00 0.17%

{100}/{112} [1�1�1], [112], [1�10] [1�1�0], [001], [1�10] 9 2

[1�1�1],

11 2

[1�1�0] 2.74×24.60×1.00 0.21%

{100}/{113} [23�3�], [311], [01�1] [01�1�], [100], [01�1] 3 2

[23�3�], 5[01�1�] 2.48×24.97×1.00 0.50%

{100}/{223} [43�3�], [322], [01�1] [01�1�], [100], [01�1] 1 2

[43�3�], 2[01�1�] 1.01×24.67×1.00 2.99%

{100}/{331} [61�1�], [133], [01�1] [01�1�], [100], [01�1] 3 2

[61�1�],

13 2

[01�1�] 3.24×25.13×1.00 0.59%

{111}/{221}(16°) [1�1�4], [221], [1�10] [1�1�2], [111], [1�10] 2[1�1�4],

7 2

[1�1�2] 3.00×25.05×1.00 1.03%

{111}/{221}(55°) [114], [221�], [1�10] [1�1�2], [111], [1�10] 2[114],

7 2

[1�1�2] 3.00×25.05×1.00 1.03%

{111}/{331}(22°) [1�1�6], [331], [1�10] [1�1�2], [111], [1�10] [1�1�6],

5 2

[1�1�2] 2.16×25.24×1.00 0.66%

{111}/{331}(49°) [116], [331�], [1�10] [1�1�2], [111], [1�10] [116], 5 2

[1�1�2] 2.16×25.24×1.00 0.66%

{111}/{100} [011�], [1�00], [011] [2�11�], [1�1�1], [011] 7 2

[011�], 2[2�11�] 1.73×23.88×1.00 1.03%

{111}/{110} [001�], [1�10], [110] [1�12�], [1�11], [110] 11[001�],

9 2

[1�12�] 3.87×24.43×1.00 0.21%

{110}/{112} [111�], [112], [11�0] [001�], [110], [11�0] 7 2

[111�], 6[001�] 2.13×24.88×1.00 1.03%

{110}/{223} [334�], [223], [11�0] [001�], [110], [11�0] [334�], 6[001�] 2.07×24.80×1.00 2.82%

{110}/{332} [113�], [332], [11�0] [001�], [110], [11�0] 3[113�], 10[001�] 3.51×26.15×1.00 0.50%

* The dimension in the y-direction includes a vacuum layer of ~4 nm.

44

Note 1. Calculation of GB energy

GB energies were calculated using periodic supercells containing a bicrystal with one internal

GB and two free surfaces. For each constituent grain, a periodic bulk reference and a slab reference

containing two free surfaces were constructed with the same orientation and relaxed in-plane

dimensions as the bicrystal. Atomic coordinates and the cell length normal to the GB were relaxed

while the in-plane dimensions remained fixed. The reference bulk energy per atom therefore

retained the elastic contribution associated with the strain imposed by in-plane periodicity. The GB

excess energy was calculated as

𝛾𝛾GB =

𝐸𝐸bi − 𝑁𝑁t 𝑒𝑒tbulk − 𝑁𝑁b 𝑒𝑒bbulk 𝐴𝐴

− 𝛾𝛾ts − 𝛾𝛾bs (S1)

Here 𝐸𝐸bi is the relaxed bicrystal energy, 𝑁𝑁t and 𝑁𝑁b are the numbers of atoms in the top and

bottom grains, and 𝑒𝑒tbulk and 𝑒𝑒bbulk are the corresponding strained bulk energies per atom. 𝐴𝐴 is

the interfacial area, equal to the area of each free surface. The surface energy of each grain 𝑖𝑖 was

calculated as

𝛾𝛾𝑖𝑖s =

𝐸𝐸𝑖𝑖slab − 𝑁𝑁𝑖𝑖slab 𝑒𝑒𝑖𝑖bulk 2𝐴𝐴

(S2)

Here 𝐸𝐸𝑖𝑖slab and 𝑁𝑁𝑖𝑖slab are the total energy and number of atoms in the two-surface slab, and 𝑒𝑒𝑖𝑖bulk is the strained bulk energy per atom. Equation (S1) isolates the GB excess energy by subtracting both the strain-matched bulk energies and the contributions from both external free

surfaces from the total bicrystal energy.

45

Note 2. Identification of SGMDs from atomistic simulations and HRTEM images

SGMDs can be identified from projected atomic-column arrangements in simulated GB

structures. This analysis provides a basis for interpreting experimental high-resolution

transmission electron microscopy (HRTEM) images and evaluating their consistency with

simulated defect structures. Atomistic simulations provide three-dimensional atomic coordinates,

whereas HRTEM images acquired along a zone axis resolve projected atomic columns in the two-

dimensional image plane. Because a single projection does not resolve atomic positions along the

beam direction, the atomic volumetric strain used to highlight dislocation cores in Fig. 3 cannot be

evaluated directly from these images. We therefore describe a procedure for identifying SGMDs

and determining their Burgers vectors through analysis of projected atomic-column arrangements.

For the four asymmetric, incommensurate GBs, {111}/{100}, {111}/{110}, {111}/{331}(22°),

and {100}/{110}, we relaxed extended bicrystal models along the horizontal direction, with free

surfaces at both ends. These boundary conditions allowed SGMDs to form during relaxation,

producing the structures shown in Fig. S4. We then traced close-packed {111} planes inclined to

the GB plane, as indicated by the white solid lines. The orientations of the terminating planes

determine the inclination angles of these {111} planes and the spacing between their intersections

with the interface. When these spacings differ between the adjoining grains, more {111} planes

intersect a given boundary length on the side with the smaller spacing. The resulting extra {111}

half-planes terminate at the interface. Yellow “⊥” symbols mark these terminations, identifying

the PGMD cores 21 . The shortest complete repeat of the PGMD arrangement defines the local GB

structural period λ, indicated by the red double-headed arrows. Crystallographic equivalence

between successive repeats was verified from the full atomic registry of the simulated structure.

Localized interruptions in the periodic PGMD arrangement identify the projected SGMD

cores, whose spans are indicated by the green double-headed arrows in Fig. S4. To determine the

Burgers vector of an SGMD, we quantified the residual mismatch within its core span relative to

the periodic PGMD arrangement in a neighboring reference span of length λ. Let 𝑁𝑁l and 𝑁𝑁u denote the numbers of projected intervals in the lower and upper grains within this reference span,

respectively. Each interval lies between successive intersections of the inclined close-packed

planes, here {111}, with the GB. The reference ratio is defined as 𝑁𝑁� = 𝑁𝑁l/𝑁𝑁u . For an SGMD core span containing 𝑛𝑛u intervals in the upper grain and 𝑛𝑛l intervals in the lower grain, the accumulated mismatch is Δ𝑁𝑁l = 𝑛𝑛l − 𝑛𝑛u 𝑁𝑁� . Positive and negative values of Δ𝑁𝑁l indicate an excess and a deficit, respectively, of lower-grain intervals relative to the reference ratio. The

46

Burgers vector is then calculated as 𝐛𝐛 = 𝐝𝐝𝐥𝐥 |Δ𝑁𝑁l|, where 𝐝𝐝𝐥𝐥 denotes the row translation vector of the lower grain along the horizontal direction. Its magnitude equals the spacing between successive

intersections of the inclined {111} planes with the GB. For the SGMDs examined here, the Burgers

vectors lie within the GB plane and perpendicular to the tilt axis, so their full magnitudes are

resolved in this projection. Atomic registry along the tilt axis was checked separately using the

near-CSL patterns in Fig. S6. Table S2 summarizes the reference and core interval counts, signed

mismatches, and corresponding Burgers vectors for all four GBs.

Fig. S5 illustrates the identification of SGMDs from projected atomic-column arrangements

in in situ HRTEM images of a {111}/{331}(22°) GB undergoing sliding. Each image in Fig. S5a,

b shows two PGMDs. Solid white lines trace extra {111} half-planes, whose interfacial

terminations identify the PGMD cores. In Fig. S5c, an additional {111} half-plane identifies an

SGMD. The associated atomic-column arrangement is consistent with the simulated structure

shown in Fig. S4c. This agreement supports the application of the procedure described in this Note

to identify SGMDs and determine their Burgers vectors from projected atomic-column

arrangements in HRTEM images.

47

Note 3. SGMD character and shear asymmetry from near-CSL analysis

Near-CSL dichromatic patterns (Fig. S6) provided an independent crystallographic check of

the Burgers vectors determined in Note 2. To construct these patterns, the two grain lattices were

superimposed in their prescribed orientations using the structural period determined from atomistic

simulation. Then, biaxial elastic strains and a relative lattice rotation were solved to produce a

common periodic cell with an exact CSL. In Fig. S6, colors distinguish the two grain lattices, while

circles and diamonds distinguish atomic layers along the ⟨110⟩ projection direction. Within each

GB plane, the shortest nonzero vector connecting lattice sites of the two grains was identified as

the minimum DSC translation within the same atomic layer. This translation connects

crystallographically equivalent interfacial states and corresponds to the Burgers vector bs of the

SGMD in the corresponding GB.

In each panel of Fig. S6, the black arrow inside the solid box denotes the shortest DSC vector

and thus the Burgers vector for the corresponding GB. Expressed in units of the lattice constant

𝑎𝑎 0 , this vector is 1/14⟨112⟩ for {111}/{100}, 1/22⟨112⟩ for {111}/{110}, 1/8⟨112⟩ for {111}/{331}(22°), and 1/24⟨110⟩ for {100}/{110}. The ⟨112⟩ and ⟨110⟩ directions are defined

relative to the grains terminated by {111} and {100} planes, respectively. These Burgers-vector

assignments agree with the projected atomic-column analysis in Note 2. In addition, each dashed

circle identifies lattice-site pairs defining the shortest DSC vectors opposite to that shown in the

corresponding solid box. These antiparallel DSC vectors have equal magnitudes and represent GB

sliding in opposite senses.

The 22° and 49° variants of the {111}/{331} GB (Figs. S6c and S8a) illustrate the respective

DSC vectors and demonstrate how atomic configurations govern direction-dependent GB shearing

responses. Both variants share the shortest DSC vector and thus Burgers vector bs = 1/8⟨112⟩{111}

and a primary period λ = 2.16 nm. For the {111}/{331}(49°) GB (Fig. S8b), red and green arrows

denote +bs and -bs, respectively. Black arrows indicate resistance from second-nearest-neighbor

atoms within each structural period beneath the terminating plane. Because this intraplanar

resistance opposes sliding along bs more strongly than along -bs, -bs represents the favored sliding

mode; conversely, sliding along bs is unfavored, which promotes shear-coupled migration for the

49° variant. In contrast, for the 22° variant, sliding is favored along both +bs and -bs, due to similar

resistance encountered in both directions.

48

Note 4. Nonuniform glide of SGMDs during pure sliding

Shear simulations of GBs containing multiple SGMDs were performed while imposing PBCs

within the boundary plane. Each supercell contained an array of SGMDs whose Burgers vectors

summed to a lattice repeat along the boundary, resulting in the formation of an extra half-plane.

This construction maintained crystallographically equivalent interfacial registry across periodic

images. The secondary periodicity vector 𝚲𝚲 connects neighboring equivalent SGMDs, and its average magnitude |𝚲𝚲�| is listed in Table 2. The horizontal supercell length therefore spans

multiple secondary periods.

For the {111}/{331}(22°) GB, sliding proceeded through intermittent glide of individual

SGMDs or subsets of the array, punctuated by collective events. Fig. S7 correlates the shear stress

response with successive dislocation positions during a representative sliding cycle. Compared

with the uniform-sliding response of the compact supercell in Fig. 4c2, this response exhibits finer

fluctuations associated with the advance of one or several SGMDs by integer multiples of the

primary structural period λ. Between states (1) and (2), the two central SGMDs advance while the

outer two remain near their previous positions (Fig. S7c1, c2). Continued loading from state (2) to

(3) raises the shear stress and produces accumulated dislocation motion along the boundary (Fig.

S7c3). This longer interval includes successive local advances during reloading. The smaller stress

decrease from state (3) to (4) accompanies further advance of three of the four tracked dislocations,

while the second dislocation from the left remains near its previous position (Fig. S7c3, c4). The

pronounced drop from state (4) to (5) corresponds to collective glide of all four dislocations (Fig.

S7c4, c5), releasing accumulated elastic shear stress during relaxation at fixed imposed strain.

49

Note 5. Cyclic loading-unloading and NEB calculations of the {111}/{110} GB

The {111}/{110} GB was loaded through 70 increments of Δγ = 2.5 × 10−4 to a shear

strain of 1.75%, and then unloaded through 70 increments of the same magnitude to zero applied

strain (Fig. 5a). Although the initial configuration was stress free, the unloaded configuration

retained a nonzero shear stress. This configuration was subsequently reloaded through 200

increments of Δγ = 0.5 × 10−4 to a shear strain of 1% (Fig. 4b2). The pronounced initial stress

drop occurred only during the first loading sequence and did not recur upon reloading. The residual

shear stress and endpoint displacements in Fig. 5a, b show that unloading did not restore the initial

interfacial structure. The NEB endpoints compare the relaxed initial and reconstructed states at the

same zero applied strain (Fig. 5c). The NEB settings and convergence criterion are specified in

Methods. The final force residual reached 2 × 10−8 eV Å−1 , below the convergence tolerance

specified there. The energy profile and saddle-point displacement field are shown in Fig. 5c, d.

50

Note 6. Shear-coupled migration

GBClaw evaluated the mechanisms governing the shear-coupled migration mode identified in

Fig. 3e using compact supercell models of the {111}/{331}(49°) and {100}/{223} GBs as

representative cases. Both exhibited pronounced stress drops associated with discrete migration

events (Fig. S9a, b). Comparing atomic configurations immediately before and after selected stress

drops linked these events to locally nonuniform atomic shuffling. Both boundaries joined an

atomically flat, low-index termination in one grain to a stepped termination in the opposing grain.

For the {111}/{331}(49°) GB, downward migration transformed atomic columns from the flat

termination of the lower grain to the stepped configuration of the upper grain (Fig. S9c1, c2). The

corresponding displacement field revealed spatially repeated, coordinated shuffling characterized

by a localized linear pattern among neighboring atomic columns (Fig. S9c3). Conversely, upward

migration of the {100}/{223} GB transformed atomic columns from the stepped termination of

the upper grain into the flatter configuration of the lower grain (Fig. S9d1, d2). Its displacement

field likewise exhibited spatially periodic but locally nonuniform shuffling (Fig. S9d3), featuring

a localized rotational pattern distinct from that of the {111}/{331} GB. In these compact supercells,

shear-coupled migration occurred collectively along the boundary rather than through the

propagation of isolated GB disconnections. Supplementary Note 7 and Fig. S10 present the

corresponding crystallographic mappings derived from dichromatic patterns. These results further

demonstrate the utility of near-CSL constructions and dichromatic patterns for determining DSC

vectors associated with distinct deformation modes in asymmetric incommensurate GBs.

51

Note 7. Determination of DSC vectors in shear-coupled migration

Using the near-CSL construction described in Note 3, we analyzed the dichromatic patterns to

identify the crystallographic mapping associated with shear-coupled migration. Atomic trajectories

yielded the signed tangential grain translation 𝑢𝑢𝑡𝑡 and normal GB displacement 𝑢𝑢𝑛𝑛 during migration of the {111}/{331}(49°) and {100}/{223} GBs. Comparing the signed ratios of 𝑢𝑢𝑡𝑡/𝑢𝑢𝑛𝑛 with the corresponding component ratios of candidate DSC vectors identified the closest DSC

vectors of 1/8⟨772⟩ and 1/4⟨5 5 12⟩ for each case, referenced to the grains terminated by {111}

and {100} planes, respectively.

Circles and diamonds in Fig. S10 distinguish atomic layers at different depths along the

common ⟨110⟩ tilt axis. Candidate connections join sites in different grains within corresponding

depth layers and may include a displacement normal to the interface. Their components were

resolved parallel and perpendicular to the dashed GB trace in the projection plane.

PBCs constrain the elastic state of the simulated bicrystals, while the near-CSL construction

also introduces strain to enforce periodic matching. These elastic states can differ, resulting in

small differences between the geometrically predicted and simulated displacement ratios.

Candidate mapping was therefore selected based on approximate agreement with the signed

displacement ratio, correspondence between atomic depth layers, and consistency with the

observed structural transformation.