Machine Learning Predicts, But Does It Explain? Causality in Biochar Research

Machine learning (ML) has rapidly become part of modern biochar research. It is now used to predict pyrolysis yield, surface area, elemental composition, adsorption capacity, greenhouse gas responses, nutrient retention, and heavy-metal immobilisation. These models are attractive because biochar performance emerges from a high-dimensional system: feedstock composition, heating history, reactor atmosphere, ash chemistry, pore architecture, surface functional groups, pollutant speciation, soil properties, and microbial processes all interact nonlinearly. However, the success of random forests, gradient boosting, support vector regression, neural networks, and other ML models has produced a new interpretability problem. Post-hoc attribution tools, including SHAP values, partial dependence plots, accumulated local effects, and permutation importance, describe how a fitted model uses variables to generate predictions. Taken alone, they do not establish ecological, geochemical, or engineering causality. In biochar studies, this distinction is critical because variables such as pyrolysis temperature, pH, ash content, specific surface area, oxygen content, and application rate are often highly collinear with one another and may act as proxies for unmeasured mechanisms. This mini review examines the attribution-causality gap in biochar research and argues that feature importance should be treated as a hypothesis-generating signal rather than mechanistic proof. The review synthesizes evidence from biochar production, pollutant adsorption, soil amendment, and nutrient cycling studies, then proposes a causal interpretability framework based on directed acyclic graphs, mechanism-aware feature engineering, controlled benchmark datasets, hybrid ML-process models, and intervention-based validation. However, this mini review is limited by its reliance on published literature rather than new experimental validation, and the proposed framework should therefore be viewed as a conceptual guide for future causal and intervention-based studies. Bridging attribution and causality is essential if ML-guided biochar design is to move from statistical optimization toward reliable, scalable, and environmentally defensible practice.