CP-Agent Automates Crystal Plasticity Simulation Workflows

Summary

Crystal plasticity simulations can predict the mechanical behavior of polycrystalline metals, but routine use often requires manually configuring heterogeneous tools, coordinating multi-step data pipelines, and calibrating constitutive parameters. The study introduces CP-Agent, a harness-engineered large language model agent that executes complete workflows from natural-language tasks. It follows the ReAct paradigm to select and sequence tools, while established optimizers perform numerical searches. Its harness uses a minimal system prompt, typed tool definitions, a dispatcher, and a safety-bounded iteration loop, with domain knowledge represented in tool schemas rather than hard-coded procedures. Four case studies demonstrate the system: calibrating four slip parameters for additively manufactured 316L stainless steel against tensile data; reproducing published copper stress-strain and texture-evolution benchmarks; recovering copper’s diffuse initial crystallographic texture; and chaining five deformation passes to reproduce the weakened, split basal texture observed in a rolled Mg-Zn-Ca alloy. Across repeated runs, the agent inferred the required execution sequence and produced physically interpretable results. The authors present harness engineering as a systematic way to automate CP workflows while retaining visible reasoning traces for auditability.