Parameter Optimization and Performance Evaluation in Polyvinyl Butyral Synthesis via Integrated Response Surface Methodology and Machine Learning

High-quality polyvinyl butyral is defined by two key performance indicators: a high degree of acetalization and small particle size. This study aims to explore the integration of response surface methodology with machine learning to optimize the synthesis process parameters of polyvinyl butyral (PVB) and facilitate model construction and evaluation for its performance indicators. Statistical analysis and advanced predictive tools can be utilized to swiftly predict and analyze product performance indicators, thereby significantly reducing costs and energy consumption. First, the response surface methodology (RSM)-central composite design (CCD) model was employed to design and optimize PVB synthesis experiments. Subsequently, RSM second-order response surface equations were utilized to fit high-quality, reliable simulation data, resulting in the construction of a support vector machine regression (SVR) machine learning model. SVR is particularly well-suited for small datasets, especially for processing data with few sample features that exhibit nonlinear relationships, showcasing good generalizability. This model demonstrated excellent accuracy, with coefficient of determination (R²) values of 0.9007 and 0.9801 for the degree of acetalization and particle size, respectively. Compared to training with the RSM model alone, the RSM-SVR hybrid model achieved significant improvements, and the RSM-CCD model with ML provided precise optimization and prediction for PVB performance indicators. Therefore, applying RSM-ML in PVB industrial production for performance analysis is of considerable value for industrial applications in materials discovery. Scientific Contribution. This study integrates RSM and SVR to optimize PVB synthesis parameters, constructing a hybrid model that achieves high prediction accuracy for acetalization degree (AD, R²=0.9007) and particle size (R²=0.9801). It innovatively uses RSM for data augmentation, reducing experimental costs by 40% while capturing nonlinear synthesis dynamics, and demonstrates that the RSM-SVR model outperforms standalone RSM or SVR models, offering significant value for industrial PVB production and materials discovery. The research highlights the synergistic advantage of combining statistical experimental design with machine learning to address the limitations of traditional kinetic models in complex phase transition processes.