A mini-review published in Journal of Machine Learning Advances by Habib Ullah, Urooj Ayaz, and Syed Sohrab Ali Shah highlights a major epistemic challenge in modern biochar research. As researchers turn to artificial intelligence to navigate the high-dimensional complexity of biomassBiomass is a complex biological organic or non-organic solid product derived from living or recently living organism and available naturally. Various types of wastes such as animal manure, waste paper, sludge and many industrial wastes are also treated as biomass because like natural biomass these More conversion, predictive algorithms have proven exceptionally effective. Advanced machine learning systems can accurately forecast biochar production yields, nitrogen content, specific surface area, heavy metal immobilization, and pollutant adsorptionBiochar has a remarkable ability to attract and hold onto pollutants, like heavy metals and organic chemicals. This makes it a valuable tool for cleaning up contaminated soil and water.   More capacity. However, the study emphasizes that high statistical prediction accuracy must not be confused with true scientific explanation.

Biochar synthesis and application represent a complex material problem where feedstockFeedstock refers to the raw organic material used to produce biochar. This can include a wide range of materials, such as wood chips, agricultural residues, and animal manure. More composition, heating regimes, chemical activation, and environmental conditions interact nonlinearly. While algorithms like random forests, extreme gradient boosting, and neural networks successfully exploit empirical patterns within tabular datasets, their internal logic reflects correlation structure rather than underlying physical laws. Consequently, popular explainable artificial intelligence tools—such as Shapley Additive Explanations, partial dependence plots, accumulated local effects, and permutation importance—describe only how a fitted model uses input variables to produce a output. They do not prove that altering a specific variable through a physical intervention will produce the expected real-world change.

This gap between model attribution and true causality arises primarily from biochar-specific sources of confounding and statistical collinearity. In biochar datasets, variables such as pyrolysisPyrolysis is a thermochemical process that converts waste biomass into bio-char, bio-oil, and pyro-gas. It offers significant advantages in waste valorization, turning low-value materials into economically valuable resources. Its versatility allows for tailored products based on operational conditions, presenting itself as a cost-effective and efficient More temperature, specific surface area, mineral ashAsh is the non-combustible inorganic residue that remains after organic matter, like wood or biomass, is completely burned. It consists mainly of minerals and is different from biochar, which is produced through incomplete combustion.   Ash Ash is the residue that remains after the complete More content, pHpH is a measure of how acidic or alkaline a substance is. A pH of 7 is neutral, while lower pH values indicate acidity and higher values indicate alkalinity. Biochars are normally alkaline and can influence soil pH, often increasing it, which can be beneficial More, and elemental ratios are strongly interdependent. When a model assigns high feature importance to pyrolysis temperature, it may simply be using temperature as a surrogate proxy for unmeasured thermal history, mineral concentrations, or feedstock choices. Similarly, treating variables measured after adsorption, such as final solution acidity, as predictive inputs introduces post-treatment bias, effectively using an outcome of the chemical process to explain the process itself.

The misinterpretation of feature attributions carries practical consequences across biochar engineering, environmental remediation, and environmental policy. In material synthesis, misidentifying specific surface area as a direct cause rather than functional group density can lead engineers to pursue expensive activation processes that offer no field-level remediation performance gain. In field applications, model recommendations developed without accounting for baseline soil context can result in inappropriate biochar dosing that causes soil salinization or nutrient imbalances. For climate certification and carbon-credit accounting, feature rankings alone cannot provide the verified proof of long-term carbon persistence required for environmental deployment.

To transform predictive machine learning into reliable scientific discovery, the authors propose a structured causal interpretability framework. Researchers should establish directed acyclic graphs before modeling to separate upstream causes, intermediate mediators, and final outcomes. Feature engineering should incorporate domain-specific knowledge, such as chemical molar ratios and pollutant speciation parameters, to reduce redundant variables. Furthermore, studies must prioritize hybrid models that embed physical mass balance and thermodynamic constraints, alongside benchmark datasets designed for controlled single-variable contrasts. Ultimately, top feature attributions must be treated as hypothesis-generating signals to be validated through targeted experimental interventions rather than accepted as mechanistic proof.

Source: Ullah, H., Ayaz, U., & Shah, S. S. A. (2026). Machine learning predicts, but does it explain? Causality in biochar research. Journal of Machine Learning Advances, 1(1), 41.