Community-Driven Innovation in Seismic Inversion Techniques

The Speed and Structure challenge aimed to advance seismic velocity inversion using AI by tasking participants with converting seismic shot records into accurate subsurface velocity models. Synthetic datasets were generated using Synthoseis and Devito, with full waveform inversion (FWI) providing benchmarks. The challenge featured 6,000 training samples, 150 test samples, and 150 holdout samples, with mean absolute percentage error (MAPE) as the key metric. Over 2,000 submissions showcased diverse approaches, many surpassing the benchmark. Top solutions employed innovative architectures and training strategies. The winning team used a Transformer-based EVA model with multi-head attention, achieving a holdout MAPE of 0.021348. The runner-up enhanced a Caformer model and expanded the training set with 20,000 synthetic samples. Third place used an ensemble of 22 EVA02 models for robust performance. Honorable mentions included Vision Transformers, hybrid loss functions, and lightweight decoders.

The challenge highlighted the effectiveness of synthetic data augmentation, architectural innovations like smoothness loss and multi-level attention fusion, and the importance of modular, well-documented code. It fostered a collaborative community and set new benchmarks for AI-driven geophysical modeling, demonstrating the potential of open-source innovation in seismic inversion.