Weak Physics‐Guided Multi‐Agent Learning for Surface to Subsurface Moisture Estimation Across Diverse Climate and Soil Conditions

Abstract Estimating subsurface soil moisture remains challenging due to limited in situ observations and the complexity of soil water dynamics. Although surface soil moisture can be retrieved from satellites with high accuracy, deeper layers are not directly observable. Traditional physics‐based models that predict subsurface soil moisture require site‐specific hydraulic properties of the soils. This limits their large‐scale applicability. Alternative data‐driven machine learning models for subsurface soil moisture estimation generally lack physical interpretability. To address the limitation of physical and machine learning models, we propose a weakly physics‐constrained, multi‐agent diffusion model for subsurface soil moisture estimation. The model employs lightweight physical regularization (flux smoothness and feasible‐range constraints) that guide predictions without enforcing strict parameterization, while a multi‐agent structure allows specialization across dry, intermediate, and wet soil regimes. This framework balances predictive flexibility with hydrological consistency and provides uncertainty quantification through stochastic diffusion sampling. The model is evaluated using globally distributed in situ data sets from 20 different sites within the International Soil Moisture Network (ISMN) and from three sites in Zambia, Africa. Soil moisture observations from ISMN are available at hourly intervals, while measurements from the Zambian stations are recorded every 10‐min. The results show a strong agreement between the modeled and observed soil moisture at multiple depths (10, 20, and 40 cm), with median values of exceeding 0.91 and nRMSE of 0.37 at 10 cm and remaining robust at deeper layers, although performance decreases with depth as expected. The model outperforms several benchmark machine learning algorithms, particularly at greater depths, and exhibits stability under stochastic initialization and input perturbations.