Available at: https://digitalcommons.calpoly.edu/theses/3130
Date of Award
6-2025
Degree Name
MS in Polymers and Coatings
Department/Program
Chemistry & Biochemistry
College
College of Science and Mathematics
Advisor
Erik D. Sapper
Advisor Department
Chemistry & Biochemistry
Advisor College
College of Science and Mathematics
Abstract
As R&D attempts to fine tune properties before production, the redesign and testing process can become highly iterative. The repetitive nature of formulating can be difficult to escape when relying on chemical intuition alone. To provide some guidance in this process, machine learning (ML) can leverage data from previous trials1 to design suggestions for subsequent trials. Although ML capabilities have vastly expanded in the last few years, integration of ML into chemistry R&D and education has been slow. Unraveling the mystique of ML can help change what has been a slow embrace in the coatings industry. It is important to demonstrate the utility of ML in industrial coatings R&D, while also developing an intuitive curriculum to prepare the next generation of chemists. By demonstrating effectiveness and ease-of-use, this paper hopes to establish the need for ML in coatings R&D and undergraduate laboratory curricula.
In this paper, different ML strategies are compared for their ability to optimize and reformulate single-component waterborne direct-to-metal (DTM) coatings. This coating system is sensitive to ingredient selection and ingredient amounts, making invalid coating formulations easy to distinguish from valid candidates in testing. Three different ML strategies are compared, built on two neural network designs. A commercial electronic lab notebook software with native ML feature integration2, is compared to two in-house models built on open-source libraries in Python. Starting with a single DTM formulation, the design space was populated with formulations by varying ingredient ratios while maintaining the ingredient list and order of addition. The ML models have a consistent problem to solve, and improvements in the model performance can be assessed over iteration of model development and material formulation. All three ML strategies struggled to predict test scores accurately, however they did manage to generate formulations that narrowed the design space towards target property scores. The neural networks struggled with making accurate predictions on a limited dataset. Combining a better model with a strong optimization strategy could still streamline the coatings R&D process at a larger scale.