Using Response Surface Methodology To Enhance Limited Data Sets for Machine Learning: A Polyvinyl Butyral Synthesis-Based Chemistry Undergraduate Teaching Case

This paper presents a novel Course-Based Undergraduate Research Experience (CURE) that addresses a critical challenge in chemical education: teaching students to generate appropriate data for machine learning (ML) applications rather than relying on precurated data sets. Through polyvinyl butyral (PVB) synthesis experiments, 22 third-year chemistry students learned to integrate Response Surface Methodology (RSM) with Support Vector Regression (SVR) algorithms to overcome the inherent data scarcity challenge in undergraduate laboratories. The course guided students through a progressive four-module framework: (1) designing experiments using Central Composite Design principles, (2) conducting PVB synthesis with varied parameters and characterizing products, (3) developing RSM models to establish mathematical relationships between synthesis parameters and material properties, and (4) using these models to generate augmented data sets for ML training. Exit survey results indicate that students self-report significant improvements in their technical competencies, particularly in experimental design (45%, n = 10), machine learning fundamentals (61%, n = 13), and data processing skills (55%, n = 12). Student-developed hybrid RSM-ML models achieved strong predictive performance (R² values range from 0.78 to 0.95 for testing data sets), significantly outperforming either methodology used independently. This integrated approach prepares students for modern research environments by developing both experimental and computational competencies. The framework is adaptable to other chemical systems.