Reliability by Architecture: Agentic Workflows for Validated Multi-Physics Simulation from Natural Language
AuthorsSøren Taverniers, Eric Darve, Hamdi A. Tchelepi
AffiliationsDepartment of Mechanical Engineering, Stanford University, Stanford, CA 94305, USA · Institute for Computational and Mathematical Engineering, Stanford University, Stanford, CA 94305, USA · Department of Energy Science & Engineering, Stanford University, Stanford, CA 94305, USA
An LLM-assisted workflow prepares and repairs multi-physics simulations, while deterministic checks help ensure the results match both the physics and the user’s request.
Key results
Grounding raised validity from 25% (5/20) to 100% (20/20).
Controlled local edits corrected all 3 seeded defects.
Tetrahedra in the demonstrated flow–poromechanics simulation.
Difference from the requested inventory, in kg.
What the paper found
This paper makes simulation reliability an architectural property: language models interpret requests, generate unsupported input components, and suggest repairs, while deterministic code constructs supported components and checks schema validity, physical admissibility, physics, and fidelity to the requested experiment. Implemented around GEOS and LangGraph, the workflow uses Anthropic’s Claude Sonnet 4.6 for deck generation and Meta’s Llama-3 8B for progress commentary; Claude Code and OpenAI Codex, using GPT-5 and GPT-6, assisted development of the tooling. In 20 flow attempts, grounding raised schema validity from 25% (5/20) to 100% (20/20), but grounding alone did not yield accepted simulations. Deterministic synthesis raised accepted yield from 0% to 100%, and normalization raised acceptance from 45% (9/20) to 100% on identical saved decks. In paired tests, controlled local edits repaired all 3 seeded input defects while leaving valid baselines unchanged. A flow–poromechanics case with frictional fault contact ran on a mesh of 22,280 tetrahedra, and its final CO2 inventory differed from the requested amount by just 0.378 kg. The central lesson is that passing a solver or conservation check is not enough: the workflow must also verify that the simulation represents the requested setup and outputs. These results demonstrate reliable automation for the tested configurations, not general accuracy or transfer to unfamiliar physics.
Original abstract
Preparing, executing, and validating strongly coupled multi-physics simulations require expert time. Agentic workflows use large language models (LLMs) to coordinate them from natural-language descriptions, but reliability requires checks of physical consistency and fidelity to the problem description. Our agentic harness implements explicit task allocation: LLMs interpret user requests and simulator diagnostics, generate input components not covered by predefined construction rules, and propose repairs, while deterministic routines construct supported components and check inputs, edits, and results. We evaluate CO2-storage problems using the multi-physics simulator GEOS, from compositional flow to flow-poromechanics with frictional fault contact. In 80 attempts on two flow problems, grounding raises the proportion of structurally valid inputs from 5/20 to 20/20, but acceptance reaches 20/20 only with complete rule-based input preparation. With identical inputs, local repairs correct three seeded defects while preserving valid inputs. Supporting tests identify simulations passing physics checks despite differing from the problem description or reference results.
Read the original paperMore in AI for Science
Browse all 58 papers →DrugCoding: A Scientific Programming Agent for Drug Discovery
Junwei Hu, Zhuohao Lin, Kun Li, Hengnan Zhang
DrugCoding is an AI agent that plans and executes multi-step computational tasks to help automate drug discovery.
Sample-Efficient Control of Magnetoconvective Flows via Differentiable Simulation
Jannis Becktepe, Dmitry Krasnov, Nils Thuerey, Jörg Schumacher, Sebastian Peitz
A differentiable simulator helps researchers control magnetoconvective flows more efficiently and design better heat-transfer outcomes.
Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
Guowei Xu, Jiaru Zou, Mete Erdogan, Haotian Ye, Pan Lu, James Zou
The paper finds that AI scientists may generate more creative ideas by first knowing less about a task, then using evidence to choose the strongest ones.