Designing biomass-derived activated carbons (ACs) is challenged by heterogeneous synthesis routes and multiobjective trade-offs among specific surface area (SBET), total pore volume (VT), and mass yield (Myield). This study presents a mechanism-enhanced multitask distillation framework (AC-ResKD) built on a shared residual deep neural network (ResDNN) and dual priors from prediction-level and teacher-aware knowledge distillation (PD-KD and TA-KD). Process factors (agent, activation temperature/time, impregnation, heating rate, precursor composition) are modeled jointly with teacher predictions to learn an interpretable mapping across combined and separated one-step/two-step datasets. All results are reported as the mean and 95% confidence intervals over 20 repeated random 80/20 splits. On separated routes, TA-KD achieves robust accuracy for SBET (one-step: R2=0.806 [0.787, 0.824]; two-step: R2=0.801 [0.773, 0.830]) and VT (one-step: R2=0.787 [0.764, 0.810]; two-step: R2=0.817 [0.791, 0.843]). On the combined set, TA-KD yields the strongest gains for Myield (R2=0.816 [0.802, 0.831]; root mean squared error ERMS=5.047 [4.853, 5.242]), while improving SBET and VT as well. Overall, Myield is the most consistent beneficiary of distillation, and route-separated training exhibits improved monotonicity and reduced bias relative to combined training; the two-step route shows stronger VT predictability consistent with a pre-carbonization priming effect. Explainability based on permutation feature importance (PFI) and Shapley additive explanations (SHAP) identifies teacher outputs, agent type, and thermal severity as dominant drivers. Impregnation governs VT in one-step activation, while pyrolysis variables rise in importance in two-step activation. Robust Pareto screening with quantile-window extraction delivers agent- and route-specific operating envelopes (temperature–dose–time), enabling simultaneous SBET/VT improvement under bounded Myield penalties. AC-ResKD thus provides accurate, interpretable, and actionable guidance for artificial intelligence (AI)-assisted AC design in heterogeneous, data-scarce settings.

Mechanism-enhanced multitask distillation for predictive, interpretable design of biomass-based activated carbons

期刊论文
发表日期：

2026年6月18日

查看全文

Peng Zhao,
Jiahui Zhou,
Guangyao Li,
Hao Yu,
Dongxu Ji,
Takahiko Miyazaki,
Kyaw Thu

摘要

Designing biomass-derived activated carbons (ACs) is challenged by heterogeneous synthesis routes and multiobjective trade-offs among specific surface area (S_BET), total pore volume (V_T), and mass yield (M_yield). This study presents a mechanism-enhanced multitask distillation framework (AC-ResKD) built on a shared residual deep neural network (ResDNN) and dual priors from prediction-level and teacher-aware knowledge distillation (PD-KD and TA-KD). Process factors (agent, activation temperature/time, impregnation, heating rate, precursor composition) are modeled jointly with teacher predictions to learn an interpretable mapping across combined and separated one-step/two-step datasets. All results are reported as the mean and 95% confidence intervals over 20 repeated random 80/20 splits. On separated routes, TA-KD achieves robust accuracy for S_BET (one-step: R²=0.806 [0.787, 0.824]; two-step: R²=0.801 [0.773, 0.830]) and V_T (one-step: R²=0.787 [0.764, 0.810]; two-step: R²=0.817 [0.791, 0.843]). On the combined set, TA-KD yields the strongest gains for M_yield (R²=0.816 [0.802, 0.831]; root mean squared error E_RMS=5.047 [4.853, 5.242]), while improving S_BET and V_T as well. Overall, M_yield is the most consistent beneficiary of distillation, and route-separated training exhibits improved monotonicity and reduced bias relative to combined training; the two-step route shows stronger V_T predictability consistent with a pre-carbonization priming effect. Explainability based on permutation feature importance (PFI) and Shapley additive explanations (SHAP) identifies teacher outputs, agent type, and thermal severity as dominant drivers. Impregnation governs V_T in one-step activation, while pyrolysis variables rise in importance in two-step activation. Robust Pareto screening with quantile-window extraction delivers agent- and route-specific operating envelopes (temperature–dose–time), enabling simultaneous S_BET/V_T improvement under bounded M_yield penalties. AC-ResKD thus provides accurate, interpretable, and actionable guidance for artificial intelligence (AI)-assisted AC design in heterogeneous, data-scarce settings.