Elucidating multi-physics interactions in suspensions for the design of polymeric dispersants: a hierarchical machine learning approach
A computational method for understanding and optimizing the properties of complex physical systems is presented using polymeric dispersants as an example. Concentrated suspensions are formulated with dispersants to tune rheological parameters, such as yield stress or viscosity, but their competing effects on solution and particle variables have made it impossible to design them based on our knowledge of the interplay of chemistry and function. Here, physical and statistical modeling are integrated into a hierarchical framework of machine learning that provides insight into sparse experimental datasets. A library of 10 polymers having similar molecular weight but incorporating different functional groups commonly found in aqueous dispersants was used as a training set in magnesium oxide slurries. The compositions of these polymers were the experimental variables that determined the complex system responses, but the method leverages knowledge of the constituent “single-physics” interactions that underlie the suspension properties. Integration of domain knowledge is shown to allow robust predictions based on orders of magnitude fewer samples in the training set compared with purely statistical methods that directly correlate dispersant chemistry with changes in rheological properties. Minimization of the resulting function for slurry yield stress resulted in the prediction of a novel dispersant that was synthesized and shown to impart similar reductions as a leading commercial dispersant but with a significantly different composition and molecular architecture.