Exascale Computational Fluid Dynamics Lab - Research Contributions

Long-Term Goals:

• Low-latency real-time inference of flow fields in High Resolution CFD.
• Accelerate existing numerical solvers.
• Submit our work to HPC & Machine Learning conferences.

Potential Approaches:

• Replacing the Runge-Kutta solver at some time-steps with a NN that predicts future flow states based on current state.
• Predict a few steps ahead and reduce total RK solves.
• Use NN to create predicted initial condition using natural language problem description.
• Train a model to learn fine-grid outputs from coarse-grid simulations.
• Upsample low-res CFD fields by training a Neural network on a ideal mapping similar to how NVIDIA DLSS works in taking a low-res render with motion vectors into a high-res image.
• Using a coase computed grid, and provided physics parameters, be able to create fine-grid solution.
• Given past time steps, model can predict next mesh without solving PDEs.
• Train model to learn how fields evolve over time, and given past ~10-20 frames, generate next frame.

Architectures Explored:

• Physics Informed Neural Networks
• Operator Learning
• CNN-Based Super Resolution
• Fourier Neural Operator For Parametric Partial Differential Equations

Tasks Completed So Far:

• Completed Standard C++ implementation of matrix operations and ported them to Kokkos for GPU acceleration.
• Read research papers and built intuition for how these CFD solvers work on deeper level, and how GPUs are leveraged and parallelized using tools like Kokkos.