This repository contains code for estimating subsurface properties—such as velocity maps—from seismic waveform data using a deep learning-based InversionNet architecture. This process, known as Full Waveform Inversion (FWI), aims to make seismic analysis more accurate, interpretable, and efficient for applications in geophysics, energy exploration, and Earth subsurface modeling.
Traditional FWI methods are computationally expensive and sensitive to noise. In contrast, InversionNet leverages the power of convolutional neural networks to directly learn the mapping from seismic waveforms to subsurface velocity models in a fully data-driven manner.
The model takes seismic waveform inputs of shape (1, 5, 1000, 70) representing:
- 5 receivers
- 1000 time samples
- 70 spatial positions
The ground-truth velocity maps have shape (1, 1, 70, 70).
You can use any synthetic dataset (e.g., generated from acoustic simulations) or real field data in compatible format.
The model is based on the original InversionNet architecture:
- Encoder: CNN layers to extract hierarchical features from seismic input
- Decoder: Transposed convolutions to reconstruct the velocity map
- Optional: Skip connections, attention layers, or temporal encoders like ConvLSTM