Development of a machine learning framework for automated geophysical prospecting using multi-physics data fusion (seismic, gravity, magnetic, and electromagnetic) to enhance subsurface characterisation.

Project Abstract

This study presents a novel machine learning framework designed to automate geophysical prospecting by fusing multi-physics data—seismic, gravity, magnetic, and electromagnetic—to improve subsurface characterisation. The framework integrates heterogeneous datasets acquired from diverse sensing modalities, addressing challenges of data alignment, resolution disparity, and noise through a unified learning architecture that combines deep learning and physics-informed constraints. A multi-branch encoder-decoder model is employed to extract complementary features from each data type, followed by a cross-modal fusion module that leverages attention mechanisms to weight modality-specific contributions based on station location, data quality, and geological context. To ensure physical plausibility, we incorporate forward-modeling constraints and regularisation terms derived from Maxwell’s equations and wave propagation theory, enabling the model to respect fundamental relationships between measured fields and subsurface properties such as velocity, density, conductivity, and magnetic susceptibility. The training pipeline utilises a hybrid loss function comprising reconstruction error for each modality, cross-modal consistency losses, and physics-based penalties that penalise non-physical predictions. A large synthetic dataset generated via high-fidelity forward simulations validates the framework, while transfer learning and domain adaptation techniques are employed to bridge the gap to field data characterized by limited labeling. The experimental design encompasses ablation studies to quantify the contribution of each modality, data fusion strategies, and the impact of varying acquisition geometries on model performance. Validation on benchmark geophysical datasets demonstrates superior subsurface characterisation compared to single-modality approaches and conventional inversion workflows, achieving higher resolution delineation of stratigraphy, fault zones, and anomalous bodies with reduced non?uniqueness. The framework also provides probabilistic uncertainty estimates through Bayesian neural components, enabling robust decision-making under data scarcity and measurement noise. A key outcome is a scalable, end-to-end toolchain that automates data preprocessing, alignment, feature extraction, fusion, and inversion-free interpretation, thereby accelerating exploration workflows and reducing reliance on expert-driven manual interpretation. In addition, the approach supports scenario testing for risk assessment, enabling rapid evaluation of potential drilling targets under varying geological hypotheses. By demonstrating improved reliability and interpretability of multi-physics integrations, this work offers a pathway toward real-time or near-real-time subsurface monitoring and prospecting in complex environments, including sedimentary basins and crystalline terrains. The study also outlines guidelines for deploying the framework in operational settings, addressing data management, computational efficiency, and reproducibility, with emphasis on extending the methodology to additional sensing modalities such as time-domain electromagnetics and gravity gradient data.

Project Overview

What This Project Is About

---

The project looks at using smart computer techniques to combine different geophysical measurements so we can better understand what lies underground. It blends data from sound waves (seismic), gravity, magnetic fields, and electrical properties to create a clearer picture of subsurface rocks and structures. The goal is to make it easier to locate resources or assess geology with less guesswork.

---

---

The Problem It Addresses

---

---

---

Objectives of the Project

---

1. Learn the basic geophysics behind seismic, gravity, magnetic, and electromagnetic methods.
2. Develop a simple framework that merges data from these methods.
3. Train a basic machine learning model to interpret fused data.
4. Test the framework on sample datasets and compare with traditional methods.
5. Assess how fusion improves accuracy and reduces misinterpretation.

---

---

What You Will Do Step by Step

---

1. Review introductory geophysics and data types. 2. Collect or simulate multi-physics datasets. 3. Preprocess data for compatibility. 4. Build a simple fusion approach (e.g., combining features). 5. Train and evaluate a basic model. 6. Visualize results and interpret subsurface features. 7. Write a short report summarizing findings. 8. Reflect on limitations and potential improvements.

---

---

Expected Outcome