AI-based Soil Moisture Solver: A Mass-Conserving Perceptron Approach

Mohammad A. Farmani, Hoshin Gupta, Ali Behrangi, and Guo-Yue Niu

Calibrating the Noah-MP land surface model poses significant computational and methodological challenges, particularly when its Fortran-based implementation relies on parameter estimates that may not be well-suited to specific regions or high-resolution datasets. Traditional calibration approaches—such as manual tuning or trial-and-error—are time-intensive and often fail to account for the complex interactions among hydrological and atmospheric processes, including soil water retention, infiltration, and baseflow generation. Furthermore, these methods do not scale efficiently in the face of rapidly expanding data availability. The reliance on Fortran exacerbates these problems by limiting the integration of modern machine learning (ML) tools, which require flexible frameworks, GPU acceleration, and differentiable programming for more efficient parameter optimization.

To overcome these constraints, we propose a conceptual surrogate model that replicates the soil moisture solver in Noah-MP while substantially reducing computational overhead. This surrogate model distills key processes—soil water retention, infiltration, and baseflow generation—into a simplified yet physically faithful representation. By eschewing the full complexity of the Fortran code, it offers a more adaptable platform for parameter optimization and hypothesis testing. Crucially, Being based on the recently developed Mass-Conserving Perceptron (MCP) concept, the surrogate model is designed to be fully compatible with contemporary ML libraries, enabling automated calibration workflows that leverage gradient-based optimization and deep learning algorithms.

By integrating advanced computational methods into a simplified yet robust modeling framework, this research aims to bridge the gap between legacy hydrological models and next-generation calibration strategies. The resulting synergy between physics-based processes and data-driven optimization promises more accurate, scalable, and efficient simulations for water resource management in arid and semiarid regions—and beyond.