GeoPINNFormer: Source Code and Reproducibility Materials
Description
This repository contains the source code and reproducibility materials accompanying the manuscript “GeoPINNFormer: Target-conditioned physics-guided spatiotemporal learning for heterogeneous hydrogeological sensor networks.” The materials implement the target-conditioned local-topology construction, event-driven temporal condensation, Transformer-based spatiotemporal encoder, discrete multi-horizon forecasting branch, continuous-query physics-guided regularization, ablation experiments, and progressive stuck-at sensor-masking evaluation. Configuration files, dependency information, example inputs, and instructions for reproducing the principal computational workflow are included.
Files
Steps to reproduce
The materials in this repository accompany the GeoPINNFormer study on target-conditioned physics-guided spatiotemporal forecasting for heterogeneous hydrogeological sensor networks. To reproduce the principal computational workflow: 1.Download all repository files and install the software dependencies specified in the accompanying README and environment/requirements file. 2.Prepare the monitoring data according to the preprocessing instructions. The observations are aligned to a daily timeline and organized into sensor metadata, historical observations/forecast targets, and environmental forcing inputs. 3.Construct target-conditioned local samples using a 30-day historical window. Each sample contains the prediction target together with its observation-supported neighboring sensors. 4.Apply event-driven temporal condensation to retain 16 temporal indices representing transient events, background states, and recent observations. 5.Train GeoPINNFormer for 150 epochs using Adam with an initial learning rate of 5×10−4. A cosine-annealing scheduler with T_max = 300 and a minimum learning rate of 1×10−6 is used. The Transformer encoder uses d_model = 64, two encoder layers, and four attention heads. 6.Evaluate the trained model using chronological rolling prediction without temporal shuffling. The discrete forecasting branch produces the reported test predictions; the continuous query branch is used only for training-time physics-guided regularization and residual diagnostics. 7.Reproduce the ablation experiments by removing event-driven temporal condensation, spatial context, or current-state anchoring individually while retaining the common training schedule. 8.Reproduce the sensor-fault robustness analysis by progressively applying zero-order-hold stuck-at masking to dynamically active pore-pressure sensors and evaluating healthy-node resilience and failed-node virtual sensing. Detailed file descriptions, execution instructions, software dependencies, and expected outputs are provided in the accompanying README.
Institutions
- Beijing University of TechnologyBeijing, Beijing