Data-Driven Approach to Tailoring Mechanical Properties of a Soft Material

Data-driven, machine learning (ML)-assisted approaches have been used to study structure-property relationships at the atomic scale, which have greatly accelerated the screening process and new material discovery. However, such approaches are not easily applicable to modulating material properties of a soft material in a laboratory with specific ingredients. Moreover, it is desirable to relate material properties directly to the experimental recipes. Herein, a data-driven approach to tailoring mechanical properties of a soft material is demonstrated using ML-assisted predictions of mechanical properties based on experimental synthetic recipes. Polyurethane (PU) elastomer is used as a model soft material to demonstrate the approach and experimentally varied mechanical properties of the PU elastomer by modulating the mixing ratio between components of the elastomer. Twenty-five experimental conditions are selected based on the design of experiment and use those data points to train a linear regression model. The resulting model takes desired mechanical properties as input and returns synthetic recipes of a soft material, which is subsequently validated by experiments. Lastly, the prediction accuracies of different machine learning algorithms is compared. It is believed that the approach is widely applicable to other material systems to establish experimental conditions and material property relationships for soft materials.