Luminary turns physics models continual: on a 30-simulation budget, car drag error falls from 10.4% to 1.6%
Synopsis
Luminary's Fall 2026 platform release replaces static large physics models with Continual Large Physics Models (CLPMs), in which agents find where a model is uncertain, run the simulations that close those gaps, and fold the new data in without forgetting what the model already knew. On a DrivAerML car aerodynamics case with a 30-simulation budget and 5 held-out test designs, the first loop's 18 evenly spread simulations left 10.4% drag error; two later loops of 6 simulations each, sent where the model was least sure, brought it to 4.8% and then 1.6%, and the loop stopped once error fell below target. The same release adds an agent-led workflow driven by plain-English requests across geometry, simulation, training and analysis.
Interpretation
Physics models no longer freeze at release: every model built on the platform ships as a Continual Large Physics Model whose capability keeps improving over its lifetime. Large physics models have been static, reproducible on a stable design space but unable to adapt when requirements change, a design review adds a parameter, or a shape outside the training data appears; updating one meant rerunning the manual cycle of simulation, training, validation and release. On a DrivAerML car aerodynamics model with a 30-simulation budget and 5 held-out test designs, the first loop's 18 evenly spread simulations gave 10.4% drag error and 17.1% lift error; a second loop of 6 simulations aimed at the least certain regions cut these to 4.8% and 7.9%, and a third loop of 6 more reached 1.6% and 4.0%, after which the error fell below target and the loop stopped on its own.
Agents carry a single plain-English request through a complete workflow spanning geometry, simulation, training and inference. In design studies the solver is rarely the bottleneck; most time goes to setup and cleanup around the solve, depends on scarce expertise, and every manual handoff creates a chance for mistakes. In the platform's examples, an agent produced a parametric geometry model ready to generate a design space in 27 minutes, turned manufacturing CAD into simulation-ready geometry in 21 minutes, built a 3D CAD model from an engineering drawing in 18 minutes, and set up and ran a full SUV aero campaign using automotive DDES best practices in 29 minutes.
Perspective
The demonstrated scope is Luminary's own solvers, data lake and model registry; the aerodynamics loop result comes from a DrivAerML car case, and the agent timing examples come from specific requests listed by the platform; the release is planned for general availability in Q1 FY2027.
The drag-error reduction comes from a single DrivAerML case with a fixed budget setting, and the material does not show whether other physics or design spaces converge the same way; the agent timing examples are records of individual requests without a stated distribution.
