Physics-Informed Neural Network Modeling of Saturation Dependent Soil Thermal Conductivity Curve

This dataset contains pre-trained neural network models and training data for predicting soil thermal conductivity (λ) as a function of soil properties and saturation degree. The dataset supports both traditional Multi-Layer Perceptron (MLP) and Physics-Informed Neural Network (PINN) approaches for modeling the saturation-dependent thermal conductivity of soils.It is organized into three main directories that support model training, inference, and visualization. The MLP directory contains all files related to the MLP approach, including the compiled dataset (Complied dataset.xlsx) with 1,846 measurements from 211 unique soil types, the trained PyTorch model weights (final_model.pth), feature scaling parameters (scaler.pkl), train-test split information (train_test_datasets.xlsx), model predictions (final_mlp_predictions.xlsx), training loss history (final_loss_history.csv), hyperparameter optimization results (hyperparameter_tuning_results.csv), and the complete training script (MLP-sample.py). The PINN directory mirrors this structure but contains files specific to the Physics-Informed Neural Network approach, with identical file types that incorporate physical constraints into the neural network architecture. Both models utilize a 75/25 train-test split by soil type to ensure generalization to unseen soils. The Load model directory provides ready-to-use inference tools, including prediction scripts for both PINN (PINN-prediction.py) and MLP (MLP-predict.py) models, along with an input data template (input_data.xlsx) that demonstrates the required format for making predictions on new soil samples. Additionally, a Chart directory contains python scripts used to generate 10 figures for the associated research publication (manuscript number: 2026WR043461), including plots of thermal conductivity curves, dataset distributions, model architecture diagrams, training results, derivative analyses, and validation comparisons. All model files are saved in PyTorch format (.pth), data files are provided as Excel spreadsheets (.xlsx) for accessibility, and the feature scalers are serialized using Python's joblib library (.pkl) to ensure consistent preprocessing during inference.