High-throughput and explainable machine learning for lacquer formulations: Enhancing coating development by interpretable models

High-throughput equipment was employed to conduct accelerated testing within a fully automated workflow, encompassing the entire process from lacquer production using raw materials to characterization through sensor-based measurements. Rather than relying on traditional design-of-experiments (DoE) for material screening, machine learning algorithms were implemented for sequential experimental design, enabling a more dynamic and data-driven approach. Specifically, a novel sampling algorithm was developed, leveraging machine learning techniques, including Gaussian process modeling and Bayesian optimization, combined with a newly designed acquisition function that iteratively calculates samples for model optimization. The effectiveness of this sampling method was validated through a toy dataset using simulated data before its application to the high-throughput formulation of a two-component polyurethane coating. In modeling the polyurethane coating, the target characteristics were gloss, hiding power, and mechanical properties such as scratch hardness and flexibility, each influenced by the various raw materials in the lacquer composition. The advantages of the machine learning model were highlighted in the post-analysis, demonstrating its capabilities in film property prediction, multi-objective modeling, and explainable artificial intelligence, all contributing to a deeper chemical understanding of coating behavior. In particular, the analysis of mechanical properties identified critical effects of the input parameters: it was found that the flexibility of the coating, within the experimental space studied, was more significantly impacted by the content of cycloaliphatic polyisocyanates than by crosslink density. Furthermore, scratch resistance exhibited complex behavior influenced by multiple factors, including hardness, which is affected by cycloaliphatic polyisocyanates and the crosslinking ratio, as well as sliding behavior influenced by the waxed matting agent. These effects were discovered with the help of explainable artificial intelligence (XAI). These findings emphasize the potential of machine learning models not only to optimize material formulations efficiently but also to provide insights into the chemical relationships underlying performance attributes, offering a path toward rational coating design. • A high-throughput lab was combined with machine learning for sequential sampling in coating formulation • A new sampling strategy Differential Area Sampling Technique (DAST) enhances data efficiency for model improvement • DAST was tested in comparison to the standard methods MaxUncertainty and Latin Hypercube Sampling • Explainable AI revealed coating behaviors in a 2-component PUR system, describing coating phenomena such as scratch resistance