Agentic harness pairs LLMs with deterministic checks for multi-physics simulation
A research paper describes a workflow in which large language models prepare and repair coupled simulations, while rule-based routines validate every input and result.
In plain words
Running simulations that combine several physical processes usually needs expert effort. The paper describes a system where language models handle interpretation and repairs, while fixed rule-based routines build and check parts of the work. In tests on carbon-dioxide storage problems, this split raised the share of usable simulation inputs.
What was announced
The authors present an agentic harness that divides work explicitly between large language models and deterministic routines. The LLMs interpret user requests and simulator diagnostics, generate input components not covered by predefined construction rules, and propose repairs. Deterministic routines construct supported components and check inputs, edits, and results, which the authors say is needed for reliability because language models alone cannot guarantee physical consistency or fidelity to the problem description.
Evaluation used CO2-storage problems on the GEOS multi-physics simulator, ranging from compositional flow to flow-poromechanics with frictional fault contact. Across 80 attempts on two flow problems, grounding raised structurally valid inputs from 5 of 20 to 20 of 20, but full acceptance reached 20 of 20 only when complete rule-based input preparation was in place. With identical inputs, local repairs fixed three seeded defects while preserving valid inputs. Supporting tests also found simulations that passed physics checks despite differing from the problem description or reference results, a limit the authors report themselves.
What it does not show
- Evidence is limited to the paper's abstract, so evaluation details, methods, and peer-review status cannot be verified from the supplied passages.
- Reported results cover two flow problems and 80 attempts; performance on other problem types or simulators is not established.
- The authors note some simulations passed physics checks while differing from the problem description or reference results, so the checks do not guarantee full fidelity.
For security and IT teams
- Scope
- The reported evaluation covers CO2-storage problems on the GEOS simulator, from compositional flow to flow-poromechanics with frictional fault contact.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.