In this study, a machine-learning-assisted strategy was presented to accelerate the preparation of high-surface-area activated carbon derived from waste wood particles. Traditional experimental approaches, such as full factorial designs, are recognized as time-consuming and labor-intensive; therefore, Bayesian Optimization and the Taguchi orthogonal array method were employed to improve experimental efficiency by 64%. Potassium nitrate was utilized as a novel activating agent, and the influence of three key parameters-activating agent mass ratio (ξ), heating rate (Q), and calcination temperature (Tc)-on the Brunauer–Emmett–Teller (BET) specific surface area (SBET) was systematically investigated. A Kriging model was constructed using an initial 16-run data set and was further refined through strategic refilling experiments, resulting in a reduction of the maximum mean squared error by 70.18%. The optimal preparation conditions were identified by the model as a mass ratio of 1.6, a heating rate of 20 °C/min, and a calcination temperature of 900 °C. Experimental validation under these parameters resulted in the production of high-purity activated carbon with a superior specific surface area of 2548.41 m2/g, deviating by only 8.7% from the predicted value. The good physical quality of the prepared sample is further supported by the adsorption–desorption curve plot, t-plot, and pore size distribution of the optimal case. These findings demonstrate that the integration of surrogate modeling and global optimization can effectively minimize resource consumption while maximizing the structural properties of biomass-derived sustainable materials.
TSAI et al. (Fri,) studied this question.