Machine Learning Approaches for Estimating Aquifer Hydraulic Properties from Step-Drawdown Pump Tests: A Case Study in Central Valley, California

We present a data-driven, scalable framework for estimating aquifer hydraulic conductivity by integrating step-drawdown pumping test data with well completion records using machine learning techniques. The approach applies Random Forest regression and cluster analysis to large regional datasets obtained from the California Natural Resources Agency and the Department of Water Resources Open Data platform. Specific capacity values derived from 3–72 hour pumping tests at 7,536 wells serve as the primary predictors of aquifer hydraulic behavior. The framework enables efficient processing of heterogeneous datasets and spatially continuous estimation of hydraulic conductivity at regional scales. Unlike traditional analytical pumping test methods, which are often constrained by limited data availability and computational demands, the proposed methodology provides a consistent, data-driven alternative for estimating hydraulic properties directly from pumping and well construction data. Results demonstrate that the machine learning framework yields robust and reproducible estimates of aquifer hydraulic conductivity at basin to subbasin scales. The approach supports the development and refinement of hydrogeologic conceptual models and provides Groundwater Sustainability Agencies with a practical tool for leveraging Sustainable Groundwater Management Act datasets to improve regional groundwater assessment and management decisions.