cs.AISep 25, 2026

CP-Agent: A Harness-Engineered Agent for Crystal Plasticity Simulation Workflows

Authors: Samuel Onimpa Alfred, Abhishek Kumar, Veera Sundararaghavan

Organizations: Department of Aerospace Engineering, University of Michigan, Ann Arbor, 48109, MI, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Crystal plasticity (CP) simulations predict the mechanical behavior of polycrystalline metals, yet their routine use is hindered by the manual effort of configuring heterogeneous tools, orchestrating multi-step data pipelines, and calibrating constitutive parameters against experiments. These bottlenecks impede productivity in systematic parameter studies, motivating interest in automated workflows. This study presents CP-Agent, a harness-engineered LLM-based agent that autonomously executes complete CP modeling workflows from natural-language tasks. Operating under the ReAct paradigm, the agent reasons about tool selection and sequencing while delegating numerical search to established optimizers. The harness comprises a minimal system prompt, typed tool definitions, a dispatcher, and a safety-bounded iteration loop, encoding domain knowledge through tool schemas rather than hard-coded logic. CP-Agent is demonstrated on four case studies: calibrating four slip parameters of additively manufactured stainless steel 316L against tensile data; validating the workflow against published copper benchmarks, reproducing stress-strain and texture evolution; recovering the initial crystallographic texture of copper, where the agent correctly identifies a diffuse initial texture; and reproducing the multi-pass rolling texture evolution of a Mg-Zn-Ca alloy, where the agent chains five deformation passes and recovers the experimentally observed weakened, split basal texture. In all cases, the agent inferred the correct execution sequence from the task statement, robustly across repeated runs, and delivered physically interpretable results. This work establishes harness engineering as a systematic approach to automating CP modeling workflows while maintaining physical interpretability and auditability through visible reasoning traces.

Figures & tables

Figure 1: Architecture of CP-Agent . The LLM is wrapped by four harness components: the system prompt, the tool definitions, the tool dispatcher, and the iteration loop. The user query enters at runtime; the agent executes a ReAct reasoning loop (Thought, Action, Observation) over the tool layer and returns a final response with physical interpretation.
ToolKey inputsDescription
inject_simulation_parametersorientation_type, fiber_direction, sigma_spread, and slip-system hardening parametersWrites texture and constitutive slip parameters into the microstructure generation script and the CP configuration file via regular-expression substitution. Enforces schema-defined defaults for any parameter not explicitly specified in the user query.
generate_microstructureNone (reads from injected config)Launches MATLAB in batch mode to generate a synthetic polycrystalline microstructure using MTEX and saves the grain ID map and per-grain Rodrigues orientation vectors to an HDF5 file. Also computes and saves pre-deformation pole figures.
convert_hdf5_to_prismsNone (reads HDF5 output)Reads the HDF5 microstructure file and converts it to the CP code input format. Applies Fortran column-major array ordering to preserve the MATLAB memory layout. Updates the CP configuration file with the correct voxel grid dimensions.
run_simulationNone (reads CP configuration file)Executes a CP simulation using the current configuration. Streams solver output in real time with a configurable timeout. Produces per-quadrature-point output files containing deformed grain orientations and stress-strain data.
extract_post_orientationsNone (reads CP output)Locates the final per-quadrature-point output file and extracts the post-deformation Rodrigues vector components. Saves extracted orientations for use in post-deformation texture analysis.
generate_pole_figuresNone (reads pre- and post-deformation orientations)Runs a MATLAB post-processing script that computes orientation distribution functions (ODFs) for both orientation sets and generates {100}, {110}, and {111} pole figures with matched color scales.
Table 1: Tools available to the agent for calibration and texture evolution workflow .
ToolAlgorithmWhen the agent selects this tool
run_bayesian_optimizationLatin Hypercube Sampling + Gaussian Process + Expected ImprovementDefault selection when no specific preference is stated. Most sample-efficient optimizer; preferred when the simulation budget is limited or an initial estimate of the optimum is unavailable.
run_differential_evolutionPopulation-based global search with mutation and crossoverSelected when the user specifies a broad or global parameter search, or when the query indicates uncertainty about the location of the optimum in the parameter space.
run_nelder_meadNelder-Mead simplex local optimizerSelected when the user provides an initial parameter estimate and requests fast local refinement. Converges quickly near the optimum but may fail to escape local minima.
run_random_boRandom Search followed by Bayesian OptimizationSelected when the user requests an exploratory phase before exploitation, or when a broad baseline is needed before Gaussian Process refinement.
Table 2: Optimizer tools available to the agent for the constitutive parameter calibration workflow .
FeatureCS1: CalibrationCS2: Forward ValidationCS3: Inverse ProblemCS4: Multi-pass Evolution
MaterialSS316LCopperCopperZX31 (Mg–3Zn–0.3Ca), HCP
QuestionWhat slip parameters fit the stress-strain data?Does the model reproduce a known benchmark?What initial texture produces the target deformation texture?How does the basal texture evolve over multiple rolling passes?
Unknown (solved for)Four slip parameters ( s0s_{0} , h0h_{0} , sss_{s} , nn )none (single forward run)initial texture (random vs fiber, spread)none (chained forward run)
Held fixedmicrostructure, loadingeverything (literature params)slip params, loading, grain counteverything (Table-1 params)
Target/referenceexperimental tensile curve [ 11 ]experimental compression curve [ 44 ] ; pole figure [ 44 ]pole figures [ 44 ]5-pass simulation + EBSD pole figures [ 45 ]
Methodagent selects optimizer →\rightarrow searchagent runs pipeline onceagent selects Bayesian optimizer →\rightarrow search over texturesagent chains five passes
Table 3: Case study specifications. The four workflows differ in their objective, unknowns, and simulation requirements while using the same agent harness with only configuration changes.
ParameterLowerUpperPhysical Meaning and its role
Initial slip resistance, s0s_{0}100 MPa150 MPaStress to initiate slip; controls yield stress
Initial hardening modulus, h0h_{0}800 MPa2500 MPaRate of initial strengthening; controls initial hardening rate
Saturation stress, sss_{s}350 MPa600 MPaMaximum achievable stress; controls the flow stress plateau
Power law exponent, nn13Strain rate sensitivity; controls sharpness of the hardening transition / strain rate sensitivity
Stopping: RMSE-<< 5 MPaAbsolute agreement criterion
Stopping: MAPE-<< 2 %Relative agreement criterion
Table 4: Search space and stopping criteria, enforced by the harness independently of the query .
Figure 2: Microstructure generation. (Left) Input grain statistics. (Right) Representative volume element (RVE) generated by the agent.
QuantityValueComment
Initial slip resistance, s0s_{0}121 MPaInterior of the bounds
Initial hardening modulus, h0h_{0}2399 MPa97.5 per cent of the upper bound
Saturation stress, sss_{s}542 MPa97.8 per cent of the upper bound
Power law exponent, nn2.9Interior of the bounds
RMSE4.95 MPaBelow the 5 MPa stopping threshold
MAPE2.85 %Above the 2 per cent stopping threshold
Table 5: Calibrated slip parameters for SS316L obtained by the agent-selected Bayesian optimizer.
Figure 3: Best-fit simulated stress-strain response (red solid line) versus experimental tensile data for SS316L (black markers and line) .
Figure 4: Convergence history of the agent-selected Bayesian optimization. Left: RMSE against evaluation number on a logarithmic scale, with the 5 MPa target as a dashed line and the converged point marked. Right: MAPE against evaluation number with the 2% target.
Figure 5: Simulated von Mises equivalent stress versus true strain, compared with the PRISMS-Plasticity TM reference curve of Yaghoobi et al. [ 44 ] .
Figure 6: Post-deformation {111}, {100}, and {110} pole figures after 100% compression on a common intensity scale (MRD 0 to 3.5): this study (top) and Fig. 10 of Yaghoobi et al. [ 44 ] (bottom), reproduced with permission. The compression axis (Z) is at the center of each projection.
Feature (MRD)This studyReference
Peak intensity, {100}\{100\}2.16∼\sim 2.0
Peak intensity, {110}\{110\}3.75∼\sim 3.5
Peak intensity, {111}\{111\}2.23∼\sim 2.0
Central {110}\{110\} ( ⟨110⟩\langle 110\rangle compression fiber)3.47∼\sim 3.5
Central {100}\{100\} (axial depletion)0.01∼\sim 0
Table 6: Quantitative pole-figure features of the post-deformation texture, compared with values read from the reference pole figures of Yaghoobi et al. [ 44 ] (0 to 3.5 MRD colorbar; reference values approximate, ±∼0.3\pm\sim 0.3 MRD).
ItemSpecification
Design variable 1Texture mode: random, or fiber about [100][100] , [110][110] , or [111][111]
Design variable 2Fiber spread σ\sigma , 3 to 85 degrees (larger spread gives a more diffuse fiber)
ObjectiveTexture-match loss vs target pole figures
Search methodBayesian optimization (Gaussian process + expected improvement)
Budget15 forward evaluations
FixedRate-dependent Taylor Cu parameters ( m=77m=77 ), compression to true strain ∼1.0\sim 1.0 , 400-grain equal-size RVE
Table 7: Inverse-problem design variables and search configuration.
Figure 7: Outcome of the 15-evaluation inverse search . (a) Texture-match loss of every evaluated initial texture versus its fiber spread σ\sigma , colored by fiber axis ( [100][100] red, [110][110] blue, [111][111] green); the dashed horizontal line is the loss of a random initial texture, the star marks the recovered optimum, and the two upward triangles at the top-left are the sharp fibers ( σ=3\sigma=3 degrees) whose losses lie above the vertical scale and are printed beside them. (b) The best loss found so far as the search proceeds (red line), with each evaluation shown as a grey point; the dotted vertical line separates the first four evaluations, a fixed initial design (random plus one fiber of each axis), from the remaining eleven, which are chosen by the Gaussian-process expected-improvement acquisition; the two evaluations whose losses exceed the axis are drawn as triangles at the top with their values labeled.
Initial textureSpread (deg)LossNote
[110] fiber270.262Recovered optimum (diffuse)
[110] fiber320.271Diffuse [110], low loss
[110] fiber370.343Diffuse [110], low loss
[111] fiber270.369Diffuse [111], low loss
[110] fiber460.436Diffuse [110]
[110] fiber850.551Very diffuse [110]
Table 8: Selected evaluations from the 15-simulation search, ordered by loss .
Figure 8: {111}\{111\} , {100}\{100\} , and {110}\{110\} pole figures on a common intensity scale (MRD 0 to 3.5), top to bottom: the recovered initial texture of this study (a diffuse [110][110] fiber of spread 27 degrees); its deformed texture after compression to true strain ∼1.0\sim 1.0 ; and the reference deformation texture from Yaghoobi et al. (2022), reproduced with permission. The compression axis ( ZZ ) is at the center of each projection.
Feature (MRD)This studyReference
Peak intensity, {111}\{111\}2.272.5
Peak intensity, {100}\{100\}2.182.35
Peak intensity, {110}\{110\}3.733.5
Central {110}\{110\} ( ⟨110⟩\langle 110\rangle compression fiber)3.113.3
Central {100}\{100\} (axial depletion)0.04∼\sim 0.3 (min)
Central {111}\{111\} (axial depletion)0.12∼\sim 0.3 (min)
Table 9: Quantitative pole-figure features of the recovered optimum’s deformed texture. Both columns are evaluated with the pole-figure scorer used in the inverse objective, so the simulated features sit on the same footing as the digitized reference targets [ 44 ] (values approximate, about ±\pm 0.2 MRD).
Figure 9: Evolution of the ZX31 (0001)(0001) basal texture under multi-pass plane-strain compression at 350 ∘350\,^{\circ} C . Columns: after the first pass (left) and the fifth pass (right). Rows: this study (top), reference simulation of Yaghoobi et al. [ 45 ] (middle), and rolling experiment (bottom). The normal direction is at the centre of each projection; RD is vertical (bottom) and TD horizontal (right).
PassThis studyYaghoobi et al. [ 45 ] Sim.Experiment [ 45 ]
12.101.947.3
55.665.464.6
Table 10: (0001)(0001) peak intensity (MRD) of ZX31 versus pass number.
More in this topic

Scientific Workflow Automation

Scientific workflow automation uses software, AI agents, and laboratory systems to plan, execute, validate, and reproduce research tasks.

See all 161 papers

Explore similar work

CardsList
May 15, 2026

ColPackAgent: Agent-Skill-Guided Hard-Particle Monte Carlo Workflows for Colloidal Packing

We introduce ColPackAgent, an agent framework that autonomously runs Monte Carlo simulations of colloidal packing through a Model Context Protocol (MCP) tool server and an agent skill, whether as a standalone agent or inside an existing agent system. By harnessing the MCP server and agent skill, ColPackAgent executes a structured workflow for colloidal packing simulations, which are central to studies of phase behavior, self-assembly, and materials design. Without dedicated simulation tools and workflow instructions, general-purpose Large Language Model (LLM) agents tend to describe such workflows rather than execute them reliably. The MCP server exposes a custom-built colpack Python package that wraps HOOMD-blue hard-particle Monte Carlo, and the skill encodes a four-stage workflow contract. ColPackAgent can carry out the workflow interactively with human feedback, autonomously from an end-to-end prompt, or as autoresearch following a provided program file. We demonstrate the system in different modes with several colloidal packing simulation examples such as cube particles in 3D, a binary system of disks and capsules in 2D, and the 2D hard-disk freezing transition using autoresearch. We also compare model performance on this workflow across a panel of LLMs with 17 stage-specific prompts. This benchmark provides a stage-level check of how reliably different models follow the setup, planning, and analysis workflow. Together, these results show that pairing a domain Python package with MCP tools and a portable agent skill provides a practical route for turning a simulation toolkit into an agent-assisted research workflow.

Lijie Ding, Changwoo Do

Neutron Scattering Division, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA

Agentic WorkflowsAI Agent BenchmarksTool-Using AgentsAutonomous Research Agents
AskPDF
May 12, 2026

CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation

Large language models deployed for MAPDL finite-element simulation face practical reliability challenges: without structured execution control, tool encapsulation, and fault recovery, outputs may be inconsistent and task failures are common. The Agent Harness paradigm addresses this by inserting domain-specific orchestration middleware that manages tool lifecycles, workflow state, and recovery escalation. This paper presents the architecture of CAX-Agent, a lightweight agent harness purpose-built for MAPDL automation, and empirically evaluates one of its core components -- the recovery policy.CAX-Agent organizes execution into three layers -- LLM service, agent harness, and solver backend -- with a recovery ladder that escalates from deterministic rule patching through model-driven regeneration to context enrichment and human intervention. We evaluate three recovery strategies (no_recovery, rule_only, and model_only) on 50 standard structural benchmarks with three repeated runs per strategy (450 case-runs total). Two independent human raters score task completion under blind conditions; inter-rater agreement is strong (quadratic weighted Cohen's kappa = 0.84, 96 percent of score pairs within one point). Model_only achieves the best completion rate (0.9267), task score (3.59/4), total score (9.16/10), and zero-intervention rate (0.84), outperforming rule_only (0.7733, 3.17/4, 7.03/10, 0.00) and no_recovery (0.6933, 2.74/4, 5.60/10, 0.00) with large effect sizes (Cliff's delta = 0.81-0.87). The benchmark uses deliberately simple geometries to isolate recovery-policy effects; we discuss the scope of these findings and directions for broader validation.

Chenying Lin, Yichen Hai, Yi He +3

Shanghai Ultradimension Technology Co., Ltd., Shanghai, China · College of Logistics Engineering, Shanghai Maritime University, Shanghai, 201306, China · School of Civil Aviation, Northwestern Polytechnical University, Xi’an 710072, China +3

LLM Agent OrchestrationLLM Agent EvaluationScientific Workflow AutomationLLM Agent Reliability
AskPDF
Sep 13, 2026

El Agente Potente: High-Throughput Agentic Atomistic Simulations

Foundational machine-learning interatomic potentials (MLIPs) are transforming atomistic simulations by achieving near-ab initio accuracy across large chemical spaces at a fraction of the computational cost. A central challenge in using these tools for high-throughput property calculations is translating high-level scientific intent into adaptive simulation campaigns without compromising workflow rigour. We introduce El Agente Potente, an agentic system that combines typed execution graphs with a complementary coding mode for MLIPs-driven atomistic simulations. Typed execution graphs provide structured and provenance-aware execution for standardized workflows, with large language models (LLMs) restricted to planning and routing while deterministic Python components perform scientific computation and validation. Complementing this structured execution, a coding agent constructs customized workflows for tasks requiring greater procedural flexibility while invoking existing Potente functions for supported calculations. We demonstrate El Agente Potente across computational materials discovery, molecular energy-landscape exploration, adsorption, and catalytic reaction workflows, together with systematic benchmarks of reproducibility and LLM token cost. These results establish typed execution graphs and code-based workflow construction as complementary mechanisms for agentic scientific computing, combining controlled, auditable execution with the flexibility required for customized atomistic simulations

Tsz Wai Ko, Jiaru Bai, Thomas Swanick +6

Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · 2Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada +7

AI Agents for Scientific DiscoveryScientific Workflow AutomationComputational Materials Science
AskPDF