Many of today’s simulation tools still rest on solver and language infrastructure from the 1980s; their view is that modern hardware and workloads justify rebuilding that stack from the ground up.

In real projects, they see the bottleneck shifting away from raw solver performance toward defining the right governing equations and system models in the first place.

Scientific ML is framed as combining known physics with data to identify missing terms or corrections in models, rather than replacing physics with purely data-driven black boxes.

They argue that traditional control schemes often rely on simplified “toy” models, and that modern solvers make it feasible to use higher-fidelity, sometimes stochastic models in advanced control strategies like MPC.

A key theme is embedding detailed models (e.g., electrochemistry + flight conditions for battery aircraft) directly into controllers to improve predictions such as remaining range and system limits.

Their target workflow is: build a high-level, multi-domain system model, then automatically derive reduced models and controller code from it, with changes in the high-level model propagating through to deployment.

They see current CAD/CAE/CFD/AI toolchains as fragmented, and argue that a composable language and solver ecosystem can reduce not just software silos but also silos between engineering disciplines.

For them, AI agents are only useful in engineering if they can iteratively generate, run, and refine models against simulations, creating a closed loop similar to how code assistants run and debug software.