Decoupling the etching-collapse competition in biochar activation: A domain-robust machine learning framework for...

Potassium hydroxide (KOH) activation is widely used to enhance the porosity of biomass-derived biochar, yet rational optimization remains difficult because activation outcomes depend on precursor texture and vary across literature sources. Here, 123 literature-derived records from 29 distinct literature sources were compiled to develop a domain-aware and interpretable machine-learning framework for biochar porosity engineering. A baseline-normalized target, Δlog(S_BET), was defined to quantify activation-induced specific surface area gain relative to the precursor state. Under source-grouped holdout validation, tree-based ensemble models retained relatively stable performance for Δlog(S_BET), with a median R² close to 0.8, whereas micropore-volume prediction showed weaker transferability. Shapley additive explanations (SHAP) and permutation-based analyses indicated that precursor baseline texture dominated the model response, with KOH/Char providing a controllable signal for activation severity. The response patterns support a literature-consistent trade-off between pore development and over-activation-induced degradation. Applicability-domain analysis was used to delineate locally supported and extrapolative regions of the design space, rather than to quantify predictive uncertainty. This framework provides domain-qualified guidance for matching activation severity to precursor texture and reducing empirical trial-and-error in biomass-derived porous carbon production.