Abstract
Despite the remarkable strides made by AI-driven models in modern precipitation forecasting, these black-box models cannot inherently deepen the comprehension of underlying mechanisms. To address this limitation, we propose an AI-driven knowledge discovery framework known as genetic algorithm-geographic weighted regression (GA-GWR). Our approach seeks to unveil the explicit equations that govern the intricate relationship between precipitation patterns and terrain characteristics in regions marked by complex terrain. Through this AI-driven knowledge discovery, we uncover previously undisclosed explicit equations that shed light on the connection between terrain features and precipitation patterns. These equations demonstrate remarkable accuracy when applied to precipitation data, outperforming conventional empirical models. Notably, our research reveals that the parameters within these equations are dynamic, adapting to evolving climate patterns. Ultimately, the unveiled equations have practical applications, particularly in fine-scale downscaling for precipitation predictions using low-resolution future climate data. This capability offers invaluable insights into the anticipated changes in precipitation patterns across diverse terrains under future climate scenarios, which enhances our ability to address the challenges posed by contemporary climate science.
Introduction
In climate science, researchers often face a trade-off between the interpretability of traditional empirical formulas and the accuracy of modern deep-learning models. While empirical models are physically intuitive, they frequently lack the precision required for complex terrains. Conversely, state-of-the-art AI models excel at forecasting but operate as "black boxes," providing little insight into the causal mechanisms—such as how specific mountain features influence local rainfall.
To bridge this gap, we introduce the Genetic Algorithm-Geographic Weighted Regression (GA-GWR) framework. This system moves beyond simple forecasting to discover explicit, interpretable mathematical equations that govern terrain-precipitation relationships. Our findings suggest that these relationships are not static; we identified a significant "1995 turning point" where the influence of altitude on rainfall began to shift globally, likely due to anthropogenic climate change.
Methodology
The GA-GWR framework integrates several advanced computational techniques to extract physical laws from data:
- Genetic Algorithm (GA): Functions as an evolutionary search engine, treating mathematical operators and terrain variables (altitude, relief, slope, aspect) as "genes" to evolve the most accurate and parsimonious equation.
- Geographic Weighted Regression (GWR): Accounts for spatial non-stationarity by allowing equation coefficients to vary across different geographic locations, respecting the principle that proximal geographic features have more similar influences.
- Physics-Informed Neural Networks (PINN): Used to link discovered coefficients to broader climate variables (e.g., temperature, vegetation), enabling the decoupling of terrain effects from climatic shifts.
- SHAP Analysis: Employs SHapley Additive exPlanations to interpret the PINN, identifying which climate factors most strongly drive changes in the terrain-precipitation relationship.
Data Sources
The study utilizes three primary datasets:
- APHRODITE: High-resolution (0.25°) precipitation and terrain data for Asia spanning 1951–2015.
- ERA5: Reanalysis data providing 33 climate variables, including atmospheric pressure and wind vectors.
- CMIP6: Future climate projections used to test the model's downscaling capabilities under various socioeconomic pathways (SSP126, SSP370, SSP585).
Results and Discussion
The Discovered Physical Law
The framework successfully derived a new explicit equation for the region: $$P = \omega_1 H + \omega_2 R + \omega_3 \cos(\alpha)H_y + \omega_4 R\beta H_x + b$$ where $H$ represents altitude, $R$ is relief, $\alpha$ is slope, $\beta$ is aspect, and $H_x, H_y$ are altitude gradients. This formula achieved a coefficient of determination ($R^2$) of 0.991, with relative errors under 5% across most grids, significantly outperforming human-derived baselines.
Climate-Driven Shifts and the 1995 Turning Point
A key discovery is that the relationship between terrain and precipitation is dynamic. Around 1995, a notable shift occurred: in the Himalayas, the impact of altitude on rainfall transitioned from a decreasing trend to an increasing one. Analysis via SHAP revealed that vegetation cover (Leaf Area Index) is a primary driver of these shifts, often exerting more influence than traditional meteorological factors like wind speed.
Applications in Downscaling
The discovered equations provide a robust mechanism for high-resolution downscaling. By applying these terrain-based formulas to low-resolution (1°) CMIP6 future climate models, we can generate detailed precipitation maps. Under high-emissions scenarios (SSP585), the model predicts significant drying in the northern foothills of the Himalayas due to the interaction between rising temperatures and local topography.
Conclusion
The GA-GWR framework demonstrates that AI can do more than just predict; it can discover. By unveiling the "hidden laws" of terrain-precipitation interaction, this research provides a tool for understanding how climate change is physically altering our environment. Future work will focus on applying this symbolic discovery approach to other complex systems, such as soil moisture and ecosystem dynamics, to further enhance our planetary understanding.
Summary updated