Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a thermodynamics-inspired explainable GeoAI framework that integrates statistical mechanics with graph neural networks to model spatial heterogeneity. The model conceptualizes spatial variability as a thermodynamic competition between system Burden (E) and Capacity (S), enabling the disentanglement of latent mechanisms driving spatial processes and identification of regime-dependent role reversals of predictors. Experiments on both synthetic and real-world datasets, including housing markets, mental health prevalence, and wildfire-induced PM2.5 anomalies, demonstrate the framework's ability to identify critical transitions and improve interpretability while maintaining strong predictive performance.
By framing spatial variability as a thermodynamic competition, this new GeoAI framework reveals hidden regime-dependent relationships that traditional methods miss, even diagnosing the shift into a "Burden-dominated" regime during the 2023 Canadian wildfires.
Modeling spatial heterogeneity and associated critical transitions remains a fundamental challenge in geography and environmental science. While conventional Geographically Weighted Regression (GWR) and deep learning models have improved predictive skill, they often fail to elucidate state-dependent nonlinearities where the functional roles of drivers represent opposing effects across heterogeneous domains. We introduce a thermodynamics-inspired explainable geospatial AI framework that integrates statistical mechanics with graph neural networks. By conceptualizing spatial variability as a thermodynamic competition between system Burden (E) and Capacity (S), our model disentangles the latent mechanisms driving spatial processes. Using three simulation datasets and three real-word datasets across distinct domains (housing markets, mental health prevalence, and wildfire-induced PM2.5 anomalies), we show that the new framework successfully identifies regime-dependent role reversals of predictors that standard baselines miss. Notably, the framework explicitly diagnoses the phase transition into a Burden-dominated regime during the 2023 Canadian wildfire event, distinguishing physical mechanism shifts from statistical outliers. These findings demonstrate that thermodynamic constraints can improve the interpretability of GeoAI while preserving strong predictive performance in complex spatial systems.