Deciphering Texture-Structure Interactions in Soil Hydraulic Behavior Through Interpretable Neural Networks

A fundamental challenge in soil physics is understanding how particle size distribution and structural organization jointly determine hydraulic behavior. Traditional analytical methods systematically destroy structural context to isolate "pure" texture measurements, eliminating the very relationships we seek to understand. While we know texture and structure interact, quantifying their relative contributions and functional interdependence across diverse soils remains elusive. We use interpretable machine learning as a discovery tool to disentangle texture and structure effects on soil water retention and hydraulic conductivity. Through staged training experiments, we systematically isolate texture-only predictions (sand, silt, clay) from structure-mediated modifications (bulk density, organic carbon). By freezing model components that interpret hydraulic behavior and controlling input availability during training, we extract learned representations that reveal how structural context alters the hydraulic meaning of identical particle size distributions. Our approach incorporates physical constraints while learning representations that capture functional complexity beyond what simple texture classes encode. Initial analyses suggest that structural inputs progressively reorganize texture-based patterns in the learned embedding space, with the magnitude of structural modulation varying systematically across soil types. Soils with identical particle size distributions occupy distinct functional spaces depending on bulk density and organic matter content—texture acquires hydraulic meaning only through structural context. These learned representations align with physical intuition: structural effects dominate precisely where classical pedotransfer functions show highest uncertainty. This demonstrates how interpretable AI can recover relationships eliminated by reductionist analytical protocols, transforming machine learning from a prediction tool into an instrument for scientific insight. Beyond improving hydraulic property estimation, the methodology offers a framework for investigating other soil properties where composition and organization interact to determine function—challenging us to rethink what we measure and how we interpret it.