← Back to Sessions

An End-to-end Physics-based Deep Learning Approach for Robust Seismic Inversion

Description: 

Seismic Inversion (SI) enables the estimation of several geophysical properties, such as velocity, a key marker for hydrocarbon exploration, earthquake analysis, and other geophysical tasks. Full Waveform Inversion (FWI) for velocity estimation in SI is computationally demanding, often requiring minutes to process as low as 32 shot gatherers on a 256 × 256 grid. Furthermore, its performance is dependent on a well-fine-tuned initial model, failing which the algorithm may face a cycle-skipping problem. Although techniques such as source encoding improve latency, FWI remains impractical for large datasets. Recently, Deep Learning (DL) has inspired physics-based SI models using unsupervised seismic data or supervised ground-truth data. However, existing DL methods often adapt designs from computer vision for their inverse networks, which are ill-suited for sensor-based inputs of SI. These methods also fail to address cycle skipping because the inverse network remains largely disconnected from the forward PDE process, leading to convergence issues during training.

We propose “SI-KCPNET,” a novel unsupervised method for efficient SI and Ultrasound Computed Tomography (UCT). SI-KCPNET embeds the physics of the second-order acoustic wave PDE into its design via three core modules: a velocity Crude Estimation Module (CEM) for roughly inverting the PDE and refinement using a Squeeze Excitation network, a Pressure Generation Module (PGM) for obtaining dense seismic data by integrating the pressure derivative using a Recurrent NN (RNN), and a Velocity Update Module (VUM) that iteratively refines the CEM estimates using an advanced RNN. The velocity map was processed using a k-t operator-based FD model with perfectly matched boundary layers. Differentiable physics enables end-to-end training, whereas a unique cycle consistency loss term mitigates cycle skipping and improves the convergence. Our unsupervised method outperforms conventional and supervised approaches on simulated UCT data, using only 25% of the training data required by supervised DL methods, and achieves greater computational efficiency than the FWI variants.

Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems - I

Type: Oral

Date: 4/15/2025

Presentation Time: 05:15 PM (local time)

Presenting Author: Mohammad

Student Presenter: Yes

Invited Presentation: 

Poster Number:

Authors