Aerospace engineers have long turned to faster computing to reduce simulation time, but Luminary argues that the real bottleneck may be the workflow itself. The company describes an aerospace OEM that moved from a conventional CPU-based solver to a GPU-native system, cutting individual computational fluid dynamics, or CFD, runtimes by nearly an order of magnitude. Yet building an aerodynamic database still took weeks.
The problem was scale and process. Aerodynamic databases can require thousands of simulations covering different geometries, flight conditions, angles of attack, sideslip, and control-surface positions. Even when individual simulations run faster, engineers must prepare, launch, validate, and process every case. A design change can force the entire batch process to begin again.
To change this workflow, the OEM and Luminary trained a Large Physics Model using several thousand existing simulations. The training data connected vehicle configurations and operating conditions with solver-generated pressure fields, forces, and moments. Once trained, the model could infer results for new conditions instead of solving every case from first principles.
In benchmark testing, the model maintained low-single-digit error compared with the solver. Avoiding a complete new simulation matrix saved nearly $10,000 in computing costs per iteration. More importantly, engineers could query the model as requirements changed instead of waiting for another multiweek simulation cycle.
High-fidelity simulation does not disappear. Solver and test data remain essential for training and validation. Uncertainty estimates can identify conditions outside the model’s validated range, indicating when additional simulations are necessary. Those results can then expand the model’s training data and trusted operating envelope.
The approach could also extend beyond aerodynamics into structural loads, flight controls, performance, and mission analysis. The larger shift is from repeatedly generating static simulation databases to creating reusable physics models that provide engineering teams with continuously accessible predictions.
