Optimization andInterpretability Analysis of MachineLearning Methods for ZnO Colloid Particle Size Prediction

To improve the efficiency of controlled zinc oxide (ZnO) colloidal synthesis, this work presented a machine learning-based method to accelerate ZnO colloid particle size control. Fifteen supervised learning methods were assessed using <i>R</i>², RMSE, and MAE metrics, with the average particle size of ZnO colloids serving as the goal variable. The hyperparameters of the two best-performing models, Random Forest (RF) and Gradient Boosting Decision Trees (GBDT), were optimized using Bayesian optimization. The SHapley Additive exPlanations (SHAP) method was employed to enhance the interpretability of the optimized GBDT and RF models. Furthermore, the Flask framework was employed to deploy the most effective GBDT-TPE model as a Web API for the purpose of predicting the average particle size of ZnO colloids. The results indicated that the RMSE of GBDT decreased by 12.9% (training) and 20% (testing) following optimization, while the RMSE of RF decreased by 22.8% (training) and 27.8% (testing). The <i>R</i>² of GBDT increased by 4.7% (training) and 29.6% (testing), while RF increased by 6.5% (training) and 16.5% (testing). The GBDT model, when optimized, surpassed RF in terms of total performance. Alkali concentration was the most significant predictor, followed by time and temperature, according to SHAP analysis. The particle size of ZnO colloids was also significantly influenced by the molar ratio, solvent, and zinc ion concentration. The largest relative error between experimental and projected results was less than 15%, indicating that the GBDT-TPE model was reliable in real applications.