Pre-inverted direct current resistivity at topography zero dataset

In the last two decades, learning-based approaches leveraging the rapid and expressive capabilities of neural networks have attracted increasing attention for solving inversion problems, particularly in geophysics. However, efficient and robust supervised training of neural networks requires large, diverse, and representative datasets. In geophysics, acquiring such datasets is often time-consuming, costly in terms of equipment and logistics, and difficult to generate numerically at scale while maintaining physical realism. To address this challenge, and within the framework of our study entitled”Neural Post-Processing of Physics-Based 2-D ERT Inversion”, we introduce the preInvDCR_T0 dataset (pre-inverted direct current resistivity at topography zero), a realistic database comprising more than $100{,}000$ samples. To the best of our knowledge, this represents the most comprehensive ERT dataset currently available in the literature. Each sample in the preInvDCR_T0 dataset consists of a 5-tuple including: (i) the ground-truth resistivity map; (ii) the mesh used to simulate the forward operator; (iii) the simulated data, including the acquisition sequence, potential differences, injected current, induced apparent resistivity, and measurement errors; (iv) the physics-based inverted resistivity map; and (v) the true RMSE of the physics-based inversion result with respect to the observed apparent resistivity. Among these, 88,702 samples with two acquisition configurations—Dipole–Dipole (DD) and reciprocal Wenner–Schlumberger (WS)—are used to train different neural networks, while 11,298 samples with the same acquisition configuration (WS) but different pre-inversion settings are reserved to evaluate the impact of pre-inversion quality on network performance. Additional information on how to access and use the dataset is available at the following link: [https://github.com/BerangerOVONO/Neural-Post-Processing-of-Physics-Based-2-D-ERT-Inversion].