3D Aerodynamic Field Prediction

Spatial AI for predicting dense physical fields on high-resolution 3D geometry from sparse sensor inputs.

  • |
  • Python
  • PyTorch
  • PyTorch3D / Open3D
  • Multi-GPU (DDP)

Predicting dense physical fields on high-resolution 3D geometry from sparse sensor inputs using geometric deep learning.

This work applies spatial AI to real-world engineering data — learning to infer unobserved physical fields (pressure, velocity, friction) over complex 3D surfaces and volumes from limited instrumentation. Models operate on point-cloud and mesh representations with millions of elements per sample, handling distribution shift from continually evolving geometry and operating conditions.

Key Aspects

  • Sparse-to-dense prediction — fusing sparse physical measurements with dense 3D geometry via cross-attention mechanisms to reconstruct full-field outputs.
  • Geometric encodings — positional encodings aligned to geodesic and adjacency structure, providing stability under extreme geometric variability and supporting reference-frame invariance.
  • Uncertainty estimation — quantifying prediction confidence to flag regions where model outputs should be treated with caution.
  • Scalable training — multi-GPU training on NVIDIA H200 clusters with benchmarking across hardware configurations for high-resolution 3D workloads.
  • Inference optimisation — 5× reduction in end-to-end inference time while preserving prediction quality for latency-sensitive applications.
  • Baseline reproduction — reproduced state-of-the-art 3D vision and geometric learning methods, delivering up to 10× reductions in memory and training time through architecture modifications.