3D geophysical data joint inversion with the concept of multimodal fusion

• Introduction of multimodal fusion to geophysical joint inversion: This study presents MP3D-NET, a deep learning neural network that effectively handles and combines multiple geophysical datasets in different dimensions, addressing the limitations in flexibility of traditional methods and minimizing information loss. • Enhanced performance for supervised learning inversions: MP3D-NET significantly improves the stability and flexibility of geophysical inversion, as demonstrated through synthetic dataset validation and realistic models. The inversion accuracy is improved through the inspection of data misfit and the incorporation of prior information about the geometry of the target. • A simple schematic of the MP3D-NET network structure, illustrating how the application of multi-modal fusion helps 3D geophysical data joint inversion. Traditional inversion methods often struggle to integrate data of different types, scales, and physical fields, leading to information loss and under-constrained models. This study proposes MP3D-NET, a deep learning-based multi-physics 3D joint inversion framework that brings the concept of multimodal fusion in machine learning to geophysical joint inversion. MP3D-NET enables joint inversion of any number of 1D, 2D, and 3D geophysical datasets, improving the stability and flexibility of handling multi-physical data. We validated the framework using synthetic datasets (block, ellipsoid, and tetrahedron) and further tested its applicability on a realistic Kimberlite exploration model. The results demonstrate that MP3D-NET effectively integrates diverse geophysical data, enhances subsurface geometric reconstructions, and optimizes inversion results through data misfit tests and the incorporation of prior information. The architecture of MP3D-NET is designed to process and integrate 1D, 2D, and 3D geophysical input data in parallel. Each data type is first processed through its own modality-specific channel, and then combined using a shared activation function. By leveraging the nonlinear mapping capability of deep neural networks, MP3D-NET directly predicts the probability of mineralization from geophysical data. Unlike traditional joint inversion methods that first reconstruct multiple physical properties models before geological interpretation, MP3D-NET embeds these relationships within the network, allowing direct geological interpretation without explicit physical model reconstruction. This study establishes a theoretical foundation for multi-modal learning in geophysical joint inversion, providing new perspectives for adopting the latest artificial intelligence technology into geophysical data processing.