Abstract:Biochar plays a pivotal role in enhancing anaerobic digestion (AD) for organic waste treatment and bioenergy recovery. By facilitating direct interspecies electron transfer (DIET) and buffering acidity, biochar can substantially improve system stability. However, the practical application of AD is often limited by the accumulation of volatile fatty acids (VFAs) and the high sensitivity of methanogens to environmental fluctuations. Although biochar provides a potential solution, its directional optimization remains challenging due to the complex, non-linear correlations among preparation conditions (e.g., pyrolysis temperature and heating rate), physicochemical properties, microbial community dynamics, and methanogenesis. Moreover, most machine learning (ML) studies treat the AD process as a "black box," mapping input materials directly to output yields while neglecting the microbial community as an essential intermediate process layer. Consequently, such models lack interpretability and fail to reveal how material properties regulate functional microbiota to improve performance. To bridge material science and microbial ecology with a primary focus on utilizing sorghum stalk, we developed a Hierarchical Data-Driven Machine Learning (HDML) framework that follows a strict "Preparation–Property–Microbe–Performance" logic. Using a Gradient Boosting Regression (GBR) algorithm, we compiled 258 datasets from the literature and our own experiments (in-house data accounted for 34.8%). The modeling was conducted in two stages to resolve the mechanism layer by layer. First, at the input layer, the model identified particle size (PS) and specific surface area (SSA) as the primary physical features governing methane yield; based on these criteria, a biochar variant with high methanogenic potential (C1) was initially screened. Second, to improve model accuracy, we introduced key microorganisms that were highly related to the methanogenic pathway by means of feature-importance analysis. Specifically, the relative abundances of the functional taxa Methanocelleus and Candidatus_Caldatribacterium were quantified as process-layer variables for iterative model refinement. Inclusion of this microbial process layer markedly enhanced model performance: the root mean square error (RMSE) decreased from 73.21 to 36.19, and the coefficient of determination (R2) increased from 0.85 to 0.87. Guided by the optimized model, the global optimal preparation conditions were determined as a pyrolysis temperature of 650 ℃ and a heating rate of 15 ℃/min (T650). Experimental validation showed that the methane yield in the T650 biochar system increased by 51% compared with the blank control (no biochar) and by 18.9% relative to the C1 group prepared without inclusion of the process layer. The T650 system also exhibited stronger pH buffering capacity and higher VFA degradation efficiency. These improvements are mainly attributed to a pore structure jointly optimized by the ideal PS and high SSA, which enhanced adsorption and buffering of inhibitory compounds and promoted DIET among enriched functional microorganisms. Overall, our findings demonstrate that HDML can elucidate the continuum from preparation conditions and physicochemical properties to microorganisms and methanogenesis, offering an interpretable and robust paradigm for the intelligent design of biochar. Close-