
Multiphysics analysis is essential when different physical phenomena interact, such as aerodynamic forces deforming an aircraft wing and that deformation subsequently changing airflow. Traditional coupled simulations repeatedly exchange results between specialized solvers, making the process computationally expensive and difficult to automate.
Luminary proposes large physics models, or LPMs, as a faster alternative. Trained on high-fidelity solver data, LPMs predict complete physical fields, such as surface pressure or structural displacement, directly from CAD geometry and operating conditions. Because they produce full-field results rather than individual quantities, they can replace conventional solvers within coupled simulation loops.
Luminary demonstrated the approach using the static aerostructural response of a transonic swept wing. The study examined wing twist and washout, which can reduce lift as aerodynamic loads deform the structure.
In the first test, Luminary replaced CFD with its SHIFT-Wing LPM while retaining the TACS finite element structural solver. Loads and displacements were transferred using MELD. The LPM-based workflow converged in 8–11 minutes on one Nvidia T4 GPU, compared with approximately four hours for CFD running on three Nvidia H100 GPUs. Across three wing geometries, predicted lift reduction differed from the CFD-based results by only 0.2–0.8 percentage points.
The second test replaced the structural solver with SHIFT-Wingbox, another LPM trained on about 1,400 TACS solutions. SHIFT-Wing and SHIFT-Wingbox then exchanged full-field aerodynamic and structural predictions directly. Results remained close to the solver-based reference, with tip twist within 0.04 degrees and lift reduction within 1 percentage point.
The demonstration suggests that coupled LPMs could make multiphysics analysis practical earlier in design. By reducing solver costs, meshing requirements, and dependence on separate specialist teams, Physics AI could allow engineers to evaluate coupled physical effects more frequently without abandoning validated engineering data.
