The DeepWave Consortium develops advanced machine-learning methods tailored for wave-equation–based seismic processing, imaging, and inversion. Our research combines the rigor of physical modeling with the scalability and adaptability of modern AI to advance subsurface characterization and monitoring. Our efforts focus on the following research directions.
Foundational models enable extraction of key features from large datasets, making downstream tasks easier to learn with higher accuracy and flexibility. In this regard, we have utilized Transformers and diffusion models through pertaining to learn the features of shot gathers or seismic recordings in general, and then used these learned features for downstream tasks like phase picking, denoising, NMO correction, velocity estimation, demultiple, low frequency extrapolation, and so on. These approaches aim to improve generalization capability of neural networks, and provide a holistic approach to address our processing and inversion workflows.
To address the limited availability of labeled field data, we develop self-supervised and weakly supervised training strategies that enhance generalization from synthetic to real seismic measurements. Our efforts include domain-robust representation learning, blind-spot and contrastive methods for denoising and signal separation, and scalable approaches applicable to both conventional seismic and distributed acoustic sensing (DAS). In addition, our work spans ML-guided full-waveform inversion (FWI) to help bridge the gap between synthetic and field data. The objective is to enable reliable deployment of ML models in operational environments.
We incorporate wave-equation constraints directly into neural network architectures and optimization frameworks to ensure physical consistency. Through Physics-Informed Neural Networks (PINNs), implicit neural representations, and differentiable simulation techniques, we bridge classical numerical methods with data-driven learning. This direction strengthens stability in ill-posed problems and enhances interpretability of learned models.
We develop data-driven priors that replace heuristic regularization in seismic inversion and imaging. By leveraging diffusion models and other generative approaches, we introduce statistically informed constraints in model space. These learned regularization strategies improve robustness, mitigate non-uniqueness, and promote geologically plausible solutions in high-dimensional and multi-parameter inversion settings.