Machine learning (ML) is increasingly used in soil science to predict soil properties, biogeochemical processes, water and solute transport, greenhouse gas emissions, and environmental risks across spatial and temporal scales. However, high predictive accuracy does not necessarily guarantee reliable scientific inference or robust extrapolation. Data leakage, inappropriate data splitting, limited representativeness of training data, poor uncertainty characterization, and a lack of process consistency can substantially undermine model reliability when ML is transferred across soils, sites, regions, or environmental conditions.

This session focuses on the development and application of trustworthy and process-informed machine learning for soil science. We aim to bring together researchers working at the interface of soil science, environmental modeling, hydrology, biogeochemistry, and artificial intelligence to discuss how domain knowledge and physical, chemical, and biological processes can be integrated with data-driven approaches.

We particularly welcome contributions addressing: (1) rigorous model evaluation and data-splitting strategies that match intended extrapolation targets; (2) data leakage, distribution shifts, domain adaptation, and cross-site or cross-region generalization; (3) uncertainty quantification, interpretability, explainability, and causal inference; (4) hybrid and process-informed ML integrating mechanistic models, physical constraints, or scientific knowledge; (5) physics-informed and theory-guided learning for soil processes; (6) multi-scale learning from laboratory experiments to field, regional, and global applications; and (7) reproducible benchmarks, open datasets, and emerging foundation-model approaches for soil and environmental sciences.

The session seeks to move soil ML beyond purely accuracy-driven prediction toward models that are scientifically consistent, interpretable, transferable, and trustworthy. We especially encourage studies that critically examine when and why machine-learning models succeed or fail, and how combining mechanistic understanding with modern AI can improve both prediction and scientific discovery in soil systems.