A thermodynamics-integrated physics-guided neural network for soil temperature forecasting.

Area of Science:

• Earth Science
• Climate Science
• Environmental Science

Background:

• Deep learning models struggle with long-term soil temperature forecasting due to error accumulation and lack of physical interpretability.
• Existing methods exhibit degraded performance in spatial generalization and capturing long-range temporal dependencies.

Purpose of the Study:

• To propose a novel forecasting framework, the Thermodynamic-Enhanced Physics-Informed Neural Network (TE-PINN), to overcome limitations in current deep learning soil temperature prediction models.
• To enhance the accuracy, physical consistency, and spatial generalization of soil temperature forecasts.

Main Methods:

• Developed TE-PINN framework using an LSTM backbone, integrating thermodynamic principles via Latent Thermodynamic Potential Inference (LTPI) and Multi-Pathway Physics-Guided Loss Integration (MPPGLI).
• LTPI module incorporates free-energy principles and dissipation constraints to model internal thermal dynamics.
• MPPGLI module employs a multi-path physics-guided loss formulation to minimize prediction-observation discrepancies.

Main Results:

• TE-PINN demonstrates slower performance degradation over multi-day forecast horizons compared to baseline models.
• The model maintains stable predictive performance across diverse datasets from different latitudes.
• Introducing thermodynamic priors significantly improves the accuracy and physical consistency of soil temperature forecasts.

Conclusions:

• TE-PINN offers a robust and accurate solution for long-term soil temperature forecasting by effectively integrating physical knowledge.
• The proposed framework addresses key limitations of traditional deep learning approaches in environmental modeling.
• This physics-informed approach advances the reliability of soil temperature predictions for agriculture, hydrology, and climate modeling.