Sparse Convolutional Autoencoder for Subsurface Reconstruction

A PyTorch implementation of a sparse convolutional autoencoder for geological subsurface reconstruction, integrating Vertical Electrical Sounding (VES) data with basement boundary modeling. This implementation is based on the methodology described in "AI-based geological subsurface reconstruction using sparse convolutional autoencoders"

Features

• Sparse convolutional architecture for efficient processing of irregular spatial data
• Multi-scale feature extraction with ConvNeXt-based encoder
• Dual-decoder design for simultaneous reconstruction of:
  • Primary variable (subsurface resistivity)
  • Secondary variable (VES measurements)
• Dynamic spatial sampling strategies
• Integration with ResIPy for VES forward modeling
• Transfer learning capabilities from pre-trained inverse distance models

Installation

Requirements

• Python 3.9+
• CUDA-compatible GPU with 8GB+ VRAM (recommended)
• 16GB+ RAM

Model Architecture

Encoder

• Four hierarchical stages with ConvNeXt blocks
• Progressive channel expansion: 96 → 192 → 384 → 768
• Sparse convolutions for memory-efficient processing
• Layer normalization and GELU activation

Decoders

Two parallel decoders with identical architectures but independent weights:

1. Primary Decoder: Reconstructs subsurface resistivity
2. Secondary Decoder: Reconstructs VES measurements

Each decoder features:

• Four inverse stages mirroring encoder structure
• UNet-style skip connections
• Transposed convolutions for upsampling
• ReLU6 activations and batch normalization