Improving Decision Stability in Ecological Restoration Planning Using Physics-Guided Sequential Learning

Title: Supplementary Material for "Improving Decision Stability in Ecological Restoration Planning Using Physics-Guided Sequential Learning"
Overview: This dataset contains the supplementary information, synthetic benchmark data, and core implementation code for the study titled "Improving Decision Stability in Ecological Restoration Planning Using Physics-Guided Sequential Learning". The research addresses a critical gap in environmental modeling: the contradiction between pointwise prediction accuracy (e.g., RMSE) and the stability of management strategy rankings under uncertainty.
We propose the PhyLSTM framework, which integrates the Advection-Dispersion-Reaction (ADR) equation as a soft regularization term into a Long Short-Term Memory (LSTM) network. This approach ensures physical consistency (mass conservation and reaction kinetics) while maintaining temporal memory, significantly improving decision robustness in ecological restoration planning.
Key Components:
1. Supplementary Document (4Supplementary Information.docx): Provides detailed derivations of the physics consistency loss, discretization schemes for the ADR equation, hyperparameter configurations, and a comprehensive reproducibility checklist.
2. Ranking Stability Sensitivity Index (RSSI): Includes the methodology and diagnostic tools for calculating RSSI (based on Kendall’s τ), a metric introduced to quantify decision robustness beyond traditional error metrics.
3. Synthetic Dataset: Due to the confidentiality of raw monitoring data from Henan Province, a synthetic benchmark dataset (data/synthetic_henan.csv) generated from the same statistical distribution is provided for code verification and performance reproduction.
4. Core Scripts:
train_phylstm.py: Main training script with physics-guided regularization.
eval_rssi.py: Script to compute RSSI and evaluate decision stability.
Technical Specifications:
Operating System: Ubuntu 20.04 LTS.
Environment: Python 3.8.10, PyTorch 1.10.0, CUDA 11.3 .
Hardware used for experiments: NVIDIA GeForce RTX 3090 GPU.
Main Findings: The study demonstrates that PhyLSTM improves RSSI by 19% over plain LSTM under 10% data availability, while keeping the mass balance error below 0.01. The results reveal a structural decoupling between prediction accuracy and decision stability, highlighting the importance of physics-guided constraints in building reliable environmental decision support systems.