Integration of PSPU-Net gravity inversion neural network with gravity data for enhanced 3D basement relief estimation

ABSTRACT Estimating the sediment-basement interface is critical to understanding basin evolution and its applications in energy, water resources, and seismic risk management. We develop a PSPU-Net gravity inversion (GI) network, a deep-learning approach combining the Pyramid Scene Parsing Network and U-Net, for gravity data to recover the sediment-basement interface. Training and validation involve smoothed basement models generated from random rectangles followed by filtering. We also incorporate uplifted basements and intrusions to enhance performance in complex geologic contexts. Numerical results for synthetic models determine PSPU-Net GI’s effective recovery of sediment-basement interface relief. To improve field data predictions, we implement transfer learning and normalization strategies. Transfer learning constructs a small number of additional basement models based on site-specific prior information and fine-tunes the neural network trained on the original general models. The normalization strategy provides a convenient way of harnessing depth information from seismic data and wells. We apply our framework to the gravity data from the western margin of the Pannonian Basin (Austria). The predictions from three implementations, baseline PSPU-Net GI, PSPU-Net GI + transfer learning, and PSPU-Net GI + normalization, successfully characterize the basement relief and are consistent with the results in previous publications. Compared with the prediction from the baseline PSPU-Net GI, the prediction accuracies obtained from the PSPU-Net GI implementations with the additional transfer learning and normalization components are notably improved.