Novel corrosion protective coating formulations suggested by machine learning of historic test data

Abstract Protective organic coatings are formulated products comprising ingredients that are selected, combined and processed to final state with desired functionality. Generally formulated products are developed using ‘craft’ approaches based on domain knowledge and experience. However the process can involve repeated cycles of (re-)formulation, laboratory trial and field test before novel products can be brought to market. In ‘recipe’ problems such as this, the combinations scale exponentially with the number of ingredients but the number of empirical trials that can be performed is limited to a small fraction of the possibilities. In this proof-of-principle contribution, we use a machine learning tool to inform selection of ingredients, and ingredient combinations, by mapping existing coating formulations to performance characteristics from laboratory testing. The available dataset was relatively sparse, comprising just 492 salt-spray test results on coatings containing 148 different ingredients and biased towards more effective formulations. The resulting multi-objective optimisation problem was solved using a differential evolution, heuristic-based method designed to restrict over-fitting. Initial model outcomes suggested two optimal formulations using novel ingredient combinations whose performance was then verified under standard cyclic salt-spray test conditions. Both novel formulations were shown to provide similar, or better, corrosion protective performance than the reference formulation.