Machine Learning Enabled Inverse Design of Coating Ink Formulation

Presented by Yuepeng Zhang, Argonne National Laboratory

The performance and durability of coating products are critically dependent on the microstructure of the coating layers, which is strongly influenced by coating and drying conditions, as well as by the formulation and rheological behavior of the inks used in the coating process. Despite advances in understanding the mechanisms governing ink property optimization during coating and drying, the development of new ink formulations with tailored properties remains largely empirical. This challenge arises from the complex, dynamic interactions among ink constituents and their sensitivity to mixing conditions. Moreover, optimizing ink properties is inherently multi-objective problem that requires balancing competing factors. For instance, increasing the binder content may enhance shear-thinning behavior desirable for slot-die coating technique, but it can also promote polymer migration and phase aggregation during drying if the drying process is not properly controlled. In this work, we use fuel cell catalyst-coated membranes as a use case to demonstrate how machine learning can be employed to correlate ink formulations with experimentally measured ink properties. We report ink-specific Gaussian Process Regression (GPR) models capable of predicting viscosity, aggregate size, and particle dispersion, enabling inverse ink design. To address the limitations imposed by small experimental datasets, we applied data-efficient modeling strategies and SHAP (Shapley Additive exPlanations) analysis to reduce experimental costs and uncover hidden correlations among parameters. This approach provides a pathway toward rational ink formulation and process optimization, advancing the integration of data-driven methods in coating science.