Agentic AI for Model-Based Engineering: How Dyad Ensures Correctness (juliahub.com)

JuliaHub’s Dyad introduces an agentic AI workflow for model‑based engineering that emphasizes correctness-by-construction rather than unconstrained code generation. Using AI agents embedded in a constrained, unit‑tested modeling language, Dyad guides model formulation (e.g., a friction‑brake thermal model), lets engineers iterate on physics-aware math before code is produced, and then enforces dimensional consistency, connector compatibility, and automated tests so generated models are deployment-ready. The system can detect and fix physical anomalies—running energy‑balance analyses, extending simulations, and revalidating behavior—catching the sorts of errors that have led to high‑profile failures (e.g., unit-mismatch disasters). Key technical points: agents produce symbolic/mathematical formulations and implement them in a language with built‑in unit checking and test harnesses; connectors carry physical types (thermal, rotational, electrical) and automatic connection validation prevents mismatches; connector-level conservation laws and unit tests ensure dimensional correctness; and automated calibration workflows (coast‑down tests) iterate parameters, run simulations, and document results. Dyad integrates cloud‑native infrastructure, differentiable programming and SciML, supports reusable parameterized components, and exposes a graphical, executable layer for inspection and editing. The result is faster model development with provable physical correctness—amplifying engineers’ productivity while reducing risky blind trust in general‑purpose code generation.