Atom Industries

Atom is a desktop app that lets engineers and researchers run physics simulations by describing them in plain language, and tells them whether to trust the answer. You type something like "a steel cantilever 1 m long with a 50 mm square section and a 1 kN tip load: what is the tip deflection?" Atom's AI agent drafts a spec, waits for your approval, then builds the model in COMSOL Multiphysics on your own machine and license, meshes it, solves it and reports the number. You can also open an existing model and ask the agent to compute, change or explain anything in it. What makes Atom different is that the AI never gets the last word on correctness. Twelve deterministic checks read the solver's own output: provenance, units, conformance to the spec, applied-load totals, solver settings, mesh convergence, balance and plausibility. Two of them solve the problem a second way without COMSOL: a trusted closed-form solution and an open-source finite-element solve. Only check code issues the verdict: VERIFIED, CHECKED or REFUSED. When a check fails, Atom says in plain words what is wrong, and the agent can fix it on a copy and verify again. Every number in an agent answer carries a badge with the verdict behind it, and every run produces an evidence report showing what each check measured and the tolerance it used. Why it matters: AI agents can now drive COMSOL, but on FEABench, Google's benchmark of real COMSOL problems, frontier coding agents solve only 5 of 15, and most failed runs end with the agent declaring success. Access to the solver is becoming free; confidence in the result is not. Where it stands: a working Windows app running against a real COMSOL Multiphysics 6.4 install. On our known set of 13 planted modeling errors across five benchmark models, the verifier catches all 13, with zero false alarms on 11 clean controls; a second set for steady laminar channel flow also gives 13 of 13 caught and no false alarm on 11 controls. These sets are development data designed alongside the checks, not a blind benchmark. Describe-to-model built six problem types it had not seen, each on the first try, for $0.96 in total API cost; judged blind, four came back VERIFIED, two CHECKED and none wrong. About 50 students around us use it for coursework and research. Everything runs locally: Atom Industries receives none of your data.