Data-Driven Property Prediction for Accelerated Discovery of Molecular Additives in Robust yet Chemically Deconstructable Thermoset Plastics

Novel materials discovery approaches are imperative to address time-critical challenges in materials design. Considering that discovery typically precedes and inspires applications, traditional screening methods often rely on arduous and inefficient experimental processes requiring the identification of novel molecules prior to application determination. Recent breakthroughs in data-science approaches have shown great promise for accelerating materials design cycles by leveraging pre-existing data to learn composition-processing-performance models which feed into active learning cycles in experimental lab spaces. However, non-standardized experimental acquisition strategies and sparsity in dataset size variability limits the ability to utilize data-driven methods for advances in the low-data regime. One such application with limited available data and which poses a time-critical environmental challenge includes the development of thermoset plastics with accessible pathways to deconstructability for improved end-of-life practices and possible reprocessing. More notably, the retention of desirable thermomechanical properties whilst introducing viable pathways to deconstructability remains a challenge. In this work, we propose the use of an ensemble of machine learning models with variable architectures, each trained on distinct parts of a training dataset consisting of data on deconstrucable polydicyclopentadiene (pDCPD) thermosets, to overcome the limitations of said low-data regime. We report a closed-loop discovery approach rooted in experimental synthesis, machine learning, and virtual screening to discover potential cleavable comonomer additives (CCAs) which, when introduced in low quantities in accordance with the reverse-gel-point model, enable the degradation of thermosets without compromising desirable thermomechanical properties. Moreover, our approach allows for concurrent prediction of resultant bulk materials properties, introducing tunability into the design process. In this work, we show it is possible to train accurate models on reconstructible pDCPD systems with as few as 101 data points to screen new chemistries and synthetic conditions, whilst predicting and experimentally validating glass transition temperatures within reasonable accuracy. This work lays a ripe foundation for the translation of closed-loop data-driven design cycles for the development of other reprocessable thermoset plastics with tailored thermomechanical properties.