Electrical resistivity tomography surveys, trained physics-informed neural network models and code for amortized ERT inversion along Route Regionale 707, Moroccan Middle Atlas

This deposit contains the field data, synthetic training datasets, trained network weights and analysis code supporting the article "Physics-informed neural network inversion of electrical resistivity tomography data: amortized optimization with field validation in the Moroccan Middle Atlas" (submitted to Journal of Applied Geophysics).   The study recasts electrical resistivity tomography (ERT) inversion as an amortized optimization problem. A U-Net trained with a physics-informed loss is fitted once and thereafter maps a measured Wenner pseudosection to a subsurface resistivity model in a single forward pass of about 2 ms, compared with 0.35 to 0.54 s per survey for conventional smoothness-constrained Gauss-Newton inversion.   FIELD DATA. Eight Wenner ERT surveys acquired at four sites along Route Regionale 707 and the adjacent N13 axis across the Tabular Middle Atlas, Morocco. Each site (El Hajeb, Ifrane, Azrou, Timhdit) was measured twice: in November 2022, before winter, and in February 2023, after the onset of winter recharge. All surveys used a Wenner array with 32 electrodes at 5 m spacing (155 m line length) and spacing levels a = 5 to 50 m, yielding 155 unique apparent-resistivity measurements per survey. Files are supplied in RES2DINV format with a manifest mapping each file to its original acquisition filename. The four sites span contrasting lithologies: karst-margin dolomitic limestone, doline-rich karst plateau, fractured basalt-capped dolomite, and Middle Liassic marls.   SYNTHETIC DATASETS. 256 generic model and pseudosection pairs used for pre-training, and 3,000 Morocco-calibrated pairs drawn from five geological archetypes representative of the survey corridor, used for domain-adaptation fine-tuning.   MODELS. Three PyTorch checkpoints: the pre-trained network, an intermediate fine-tuning checkpoint, and the final model used for all results reported in the article. The architecture is a U-Net with base dimension 32 and 1,927,841 trainable parameters, mapping a 24x96 log-apparent-resistivity pseudosection to a 32x64 log-resistivity model spanning 100 m horizontally and 50 m in depth.   CODE. Eight Jupyter notebooks covering synthetic data generation, pre-training, Morocco fine-tuning, field inference, the conventional PyGIMLi inversion baseline, model comparison and the forward-pseudosection physics check. A framework-independent NumPy re-implementation of the inference pipeline is included as a reproducibility check that does not require PyTorch.   Both synthetic sets were forward-modelled with the Gaussian sensitivity-kernel operator described in the article rather than with a finite-element solver. The README documents this, together with a note on the stored used_pygimli flag, which records only that the PyGIMLi import succeeded and not that PyGIMLi performed the forward modelling.