Reproduce “SubsurfaceGen: Procedural Generation of Field-Scale Earth Models and Seismic Data” on OpenResearch

About this paper

Joseph Stitt, Pratik Rathore, Madeleine Udell, Ching-Yao Lai

Researchers at Stanford University developed SubsurfaceGen, a GPU-accelerated procedural generator for field-scale 3D earth models and seismic data, creating a diverse dataset of 42 velocity models and over 21,000 paired wavefields/shot gathers. The system demonstrated up to 38.9x GPU speedups for model generation and enabled experiments revealing long-horizon error accumulation in neural operators, which was mitigated by a novel TFNO-interp method, alongside architectural differences in generalization for end-to-end inversion.

Key insights

• Generating and predicting wavefields at field scales (5-second trajectories on large grids) exceeds typical GPU memory, necessitating chunked autoregressive rollouts that expose significant long-horizon error compounding.
• The proposed TFNO-interp, an adaptation of optimal checkpointing for neural operators, effectively mitigates error accumulation in wavefield prediction by anchoring predictions between stored wavefield states.
• Different ML architectures for end-to-end inversion exhibit varying performance in generalizing to out-of-distribution geological settings, with 3D CNNs capturing large-scale features more accurately but Transformers showing more stable SSIM on OOD data.

Summary by alphaXiv.