Experimental validation of products emerging from complex chemical processes is slow and expensive, limiting throughput and constraining exploration of vast design spaces in materials discovery. Meanwhile, decades of patents and scientific reports contain detailed but heterogeneous records—spanning text, tables, and molecular structures. We present a framework that turns this fragmented measurements into actionable predictions by (i) introducing a universal, directed-tree representation of chemical processes and (ii) training a multi-modal, attention-based graph neural network (GNN) that learns transferable process embeddings for cross-domain property prediction.

In our graph-based representation, nodes capture structural containers such as materials, processes, conditions, properties, along with leaf node modalities including numeric values, text, and molecular structures. We employ state-of-the-art encoders for each modality: large language models for text descriptions, vector embeddings for numeric data, and molecular GNN for molecular structure embeddings. These heterogeneous embeddings are then fused into a single chemical process graph and processed by edge-aware transformer message passing GNN model. A property-conditioned attention mechanism identifies the most predictive materials, steps, and conditions for prediction of the queried property.

We train our GNN backbone on hundreds of thousands of processes from patents and literature across diverse chemical domains. The multi-task architecture features a shared encoder that produces universal process embeddings, paired with lightweight task-specific heads for property prediction. By unifying heterogeneous experimental evidence into process-centric graphs, our approach accelerates materials R&D by prioritizing high-value experiments and supporting both autonomous and human-in-the-loop discovery workflows.