Machine learning for adhesion assessment and prediction

Adhesion is governed by a combination of several chemical and physicochemical interactions at the coating-substrate interface. Traditional adhesion experiments are often resource-intensive and provide limited insights into the multiple factors influencing adhesion outcomes. Recent studies have applied machine learning (ML) methods to the modelling and prediction of adhesion properties, and this review examines existing ML approaches for adhesion assessment across coatings and adhesive systems. Here, published studies are analysed in terms of dataset size and quality, feature selection strategies, ML model choice, validation practices, and reported performance metrics. In doing so, we highlight common limitations, including small and heterogeneous datasets, limited use of independent test sets, and inconsistent adhesion testing protocols, which frequently lead to overestimated predictive performance. We also re-evaluate selected case studies to illustrate how data handling choices and experimental design can influence model accuracy and generalizability. Based on these observations, we identify key challenges and opportunities for data-driven adhesion modelling and provide recommendations to improve data transparency, incorporate chemically meaningful descriptors, and adopt robust validation frameworks. The review concludes by outlining future directions for integrating ML with experimental design to support more reliable and scalable adhesion prediction for advanced coating and adhesive applications.