Predicting Soil Organic Carbon Response to Biochar Using Parsimonious Machine Learning Models

Introduction. Accurate prediction of the soil organic carbon (SOC) response ratio (RR) following biochar application remains challenging due to substantial heterogeneity in environmental conditions, soil properties, and experimental designs across the published literature. Predictive models that incorporate a large number of variables often reduce the available sample pool, which limits their capacity to generalize across contexts.

Methods. This study presents a machine learning pipeline trained on a global dataset compiled from hundreds of biochar trials conducted across multiple climatic regions. A parsimonious, data-consistent feature space was defined with five mechanistically interpretable predictors: biochar application rate, initial SOC content, soil pH, climate zone, and treatment type. These variables were selected based on their consistent availability across datasets and their physicochemical relevance to carbon cycling dynamics. A suite of models, including linear, regularized, and tree-based algorithms, was trained with systematic hyperparameter optimization. Adjusted R² served as the primary performance criterion for model selection and comparison.

Results. Models trained on the five-predictor feature set achieved competitive and consistent predictive performance. Expanding the feature space beyond this parsimonious core did not produce substantial accuracy gains. On the contrary, the reduction in available observations associated with higher-dimensional configurations decreased generalization capacity.

Conclusions. A concise set of physically meaningful predictors is sufficient to reliably estimate SOC response to biochar amendment across diverse global contexts. This framework demonstrates that matching model dimensionality to data completeness is critical for robust performance. The proposed approach provides a scalable and interpretable tool for soil carbon assessment with direct applicability to Earth system modeling and land management decision support.