Formulation Development Using Machine Learning: From Analyzing Experiments to Predicting Performance

Surfactant phase behavior is a critical aspect of liquid formulation development in multiple applications ranging from detergents & cleaners to agrochemicals & enhanced oil recovery. Often these applications use multiple surfactants along with other components including solvents, salts, and polymers. Predicting the long-term phase behavior of these fully formulated systems remains challenging. These challenges include predicting when an emulsified formulation will separate and will a formulated system remain 1 phase at both elevated and sub-ambient temperatures.

Here we describe a multi-year effort to apply machine learning models to formulated systems containing surfactants to advance product and application development. We demonstrate the utility of these models in processing raw data and the modelling of the results to improve formulation stability. More specifically, we utilize high throughput experimental studies to develop datasets of formulations to model surfactant phase behavior in fully formulated systems. Analysis of the phase behavior from thousands of samples leveraged a custom machine learning image processing model to assess the system's stability and classify potential failure modes. This approach demonstrated that these models can accurately identified the phase behavior >90% of the time. Next, we demonstrate that combining these experimental data sets with physical and chemical descriptors of the formulation components enable training of ensembles of cluster-based machine learning models to predict whether a given formulation would be stable, some of which demonstrated cross-validation accuracies as high as 75%. We then demonstrate this approach on a second application with a separate dataset of >2500 samples resulting in accuracies >90% for phase stability and >85% accuracy for cloud point. Finally, we discuss the challenges necessary to extend the modelling approaches to application performance testing.