Development of a Physics-Informed Neural Network for Joint Inversion of Gravity and Magnetic Data in Complex Basement Terrain, Nigeria

Joint inversion of gravity and magnetic data using conventional regularised methods is computationally intensive and requires subjective a priori structural coupling constraints that are difficult to define systematically in geologically complex Nigerian basement terrains where both density and susceptibility vary strongly and independently in three dimensions. This research develops a physics-informed neural network architecture that performs joint inversion of gravity and magnetic data by encoding the physical forward operators as differentiable layers within the network and learning structurally coherent subsurface models directly from multi-physics data without explicit regularisation. The PINN architecture is trained on a synthetic dataset of 50,000 geologically realistic subsurface scenarios generated from stochastic Gaussian random field models parameterised with Nigerian basement complex lithological properties. The trained network is evaluated on a held-out test set and on two real-world datasets from the Basement Complex of Ekiti and Taraba states. On the synthetic test set, the PINN recovers density and susceptibility models with 12 percent lower RMS error than a conventional gradient-based joint inversion at one-hundredth of the computational cost. Application to the Ekiti dataset produces a basement lithological map that agrees with geological mapping at 78 percent of outcrop control points. The research demonstrates that deep learning physics-informed inversion can replace computationally prohibitive conventional approaches while maintaining geophysical rigour, contributing an original methodological advance to geophysical practice in data-limited exploration environments. Keywords: physics-informed neural network, joint inversion, gravity, magnetic, basement complex