Biochar is widely recognized as an effective approach for mitigating heavy metal pollution. However, the utilization of machine learning models to guide biochar preparation and enhance its adsorption performance poses challenges. This study proposed a biochar design strategy guided by the Bayesian optimization algorithm. The strategy involves automated hyperparameter optimization for machine learning model training and exploration of unexplored biochar preparation conditions using the Bayesian algorithm. By leveraging Bayesian algorithms for hyperparameter search, we successfully crafted random forest models, support vector regression models, and back propagation models. In comparison to the previously reported models, these models demonstrated outstanding performance, with the random forest model in particular showcasing superior results (R2 = 0.998; RMSE = 0.027). Using a Bayesian algorithm, it was found that more than 80% of the feature combinations would exceed the upper limit of heavy metal adsorption. Our model revealed that biochar with a mesoporous structure, produced at a pyrolysis temperature of 420 °C, exhibited enhanced heavy metal adsorption capacity. This study presented a novel approach for the rapid development of machine learning models using Bayesian optimization and employed inverse reasoning to guide biochar preparation and enhance adsorption performance.