Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1408
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-15T06:54:04.215586Z
status: published
lens: g_space
sector: science
headline: LLM-driven design of physics-constrained constitutive models: Two agents are better than one
dek: Developing constitutive models that capture how materials deform under load traditionally requires years of specialized expertise in continuum mechanics, machine learning, and scientific programming. Large language models (LLMs) have recently been shown to lower this barrier by generating constitutive models on demand, but existing single-agent pipelines lack systematic checks that the resulting models respect fundamental physical laws. To close this gap, we introduce the first multi-agent LLM-driven approach fo…
gain_title: Adding an Inspector agent that checks each proposal against nine physical constraints increased the share of generated constitutive models passing all checks, from 90% to 95% for Opus and from 47% to 60% for Kimi.
problem_title: (none)
trace_subject: (none)
gain_reading: Adding an Inspector agent that checks each proposal against nine physical constraints increased the share of generated constitutive models passing all checks, from 90% to 95% for Opus and from 47% to 60% for Kimi.
gain_evidence: Adding the Inspector raises the share of exported models that pass all these checks from 90% to 95% for Opus and from 47% to 60% for Kimi.
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers built a two-agent system for generating constitutive models, the equations describing how materials deform under load. A Creator agent proposes a constitutive artificial neural network tailored to the data, and an Inspector agent audits each proposal against nine physical constraints, sending it back for refinement when a violation is found.

This matters because single-agent LLM pipelines can produce models without systematic checks on physical validity, so separating generation from inspection makes automated modeling more trustworthy. Even with the Inspector, a considerable share of Kimi-generated models still failed at least one check, and the tests covered only a handful of materials.
limitation: Even with the Inspector, only 60% of Kimi's exported models and 95% of Opus's passed all physical checks, so some generated models still violated constraints, and testing covered only three materials.
tag: Evidence-backed gain
key_points: The system was benchmarked on brain tissue and rubber as isotropic materials, and on porcine skin as a transversely isotropic material with a fiber direction. | Two LLM backbones were tested: Claude Opus 4.7 and Kimi K2.5. | Constraint compliance was assessed numerically across a broad sample of deformation states, rotations, and perturbation directions. | Generated models matched or surpassed expert-designed models in accuracy and extrapolated beyond the training data.
rundown: Traditionally, building constitutive models takes years of expertise in continuum mechanics, machine learning, and scientific programming. Existing LLM approaches lower this barrier by generating models on demand but lack systematic physics checks.

The authors present the first multi-agent LLM-driven approach for this task. The Creator proposes a model, and the Inspector returns it for refinement whenever any of nine physical constraints is violated.

Pass rates rose from 90% to 95% for Opus and from 47% to 60% for Kimi. The generated models were reported to be on par with or better than expert-designed ones and to generalize well to unseen loading paths. The authors describe the approach as technique-agnostic and expect it to scale with LLM capability.
sources:
- peer_reviewed | Machine Learning for Computational Science and Engineering | https://doi.org/10.1007/s44379-026-00099-3 | 2026-10-01
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

6466d20da32edaf5b78782c60a8eeab20f6490311d5d44934155619eed8afc30

previous

0000000000000000000000000000000000000000000000000000000000000000