ACCELERATING MATERIALS DISCOVERY: LEARNING A UNIVERSAL REPRESENTATION OF CHEMICAL PROCESSES FOR CROSS-DOMAIN PROPERTY PREDICTION

Abstract

Abstract Experimental validation of chemical processes is slow and costly, limiting exploration in materials discovery. Machine learning can prioritize promising candidates, but existing data in patents and literature is heterogeneous and difficult to use: records combine free text, numeric conditions, molecular structures, and variable-length process steps, so fixed-length flattening loses step order and condition--material links. We introduce a universal directed-tree process-graph representation that unifies unstructured text, molecular structures, and numeric measurements into a single machine-readable format; we adopt a directed tree precisely because it preserves this causal step-and-condition structure while accommodating processes of widely varying size. To learn from this structured data, we developed a multi-modal graph neural network with a property-conditioned attention mechanism. Because any single application typically provides only a handful of labelled experiments, we pretrain on hundreds of thousands of process records extracted and curated from patents and scientific literature, enabling a shared representation to transfer to new tasks with limited training data. Trained on approximately 700,000 process graphs from over 10,000 diverse documents, our model learns semantically rich embeddings that generalize across domains. When fine-tuned on compact, domain-specific datasets, the pretrained model achieves strong performance, demonstrating that universal process representations learned at scale transfer effectively to specialized prediction tasks with minimal additional data.