Met-Shield PINN simulates re-entry thermal gradients in browser at 60fps
A PINN trained on the 3D heat equation runs natively in WebAssembly at 60 fps.
Met-Shield is a real-time re-entry simulator that leverages a Physics-Informed Neural Network (PINN) to replace traditional finite difference solvers. The PINN uses a fully connected architecture trained to minimize the residual of the 3D heat equation, constrained by the thermal diffusivity and conductivity of Ti-6Al-4V (a titanium alloy used in spacecraft). The goal is to achieve robust generalization on noisy atmospheric trajectory data, which standard solvers struggle with.
Once trained, the model is integrated into a custom C++ engine compiled to WebAssembly, enabling the simulation to run at 60fps directly in a browser. This allows on-the-fly prediction of metallurgical phase transitions (alpha-to-beta titanium) during re-entry. The developer notes convergence challenges during the Max Q phase (peak aerodynamic pressure) and is seeking advice on weighting physics-loss vs. data-loss. The project is fully open-sourced on GitHub with a live demo.
- Uses a Physics-Informed Neural Network (PINN) to solve the 3D heat equation for re-entry thermal analysis.
- Model compiled to C++/WebAssembly runs at 60fps in the browser, predicting Ti-6Al-4V phase transitions.
- Open-source project faces convergence issues during Max Q phase; developer requests feedback on loss weighting.
Why It Matters
Browser-based real-time aerospace simulation reduces reliance on supercomputers, democratizing spacecraft thermal analysis.