Data science can accelerate materials discovery by learning composition-processing-performance models from pre-existing data sets, which then feed into active learning cycles in the laboratory. Thermoset polymer waste is a pressing environmental challenge that may be addressed by the accelerated discovery of new deconstructable variants; however, the combinatorial space of possible monomers, crosslinkers, additives, and manufacturing conditions is vast and Edisonian experimentation may struggle to find optimal designs. Moreover, data-driven strategies are limited for complex (co)polymers like thermosets because the training data is scarce and sourced from heterogeneous experimental approaches, resulting in overfit transferable models. Here, we introduce a novel closed-loop approach to the predictive design of chemically deconstructable thermosets that leverages experimental synthesis and characterization, machine learning, and virtual screening. Our computational model learns to map the molecular features of bifunctional silyl ether (BSE)-based cleavable comonomers to the thermal properties of the industrial thermoset polydicyclopentadiene (pDCPD). We address the challenges of limited data and overfitting by relying on both structural and information-rich domain-specific molecular features as inputs and by thoroughly quantifying model uncertainty. By training an ensemble of predictive models mixing multiple model architectures and parametrizations, our approach achieves predictions of a key thermoset parameter—the glass transition temperature—within less than 15 °C error over a wide temperature range with only 101 data points. The trained models were used to screen new possible BSE comonomer compositions and synthesis conditions, with promising combinations successfully validated experimentally. This work offers a closed-loop design process that we expect to be widely applicable to the discovery of deconstructable polymeric materials.
Supplementary Information for A Model Ensemble Approach Enables Data - Driven Property Prediction for Chemically Deconstructable Thermosets in the Low Data Regime