Machine learning-driven molecular generation for accelerated screening of high-performance flame retardants in epoxy resin composites

The development of flame retardants (FRs) for epoxy resins (EPs) has historically relied on empirical trial-and-error approaches. Herein, we present a novel Machine Learning (ML) framework integrating unsupervised learning, supervised learning, and molecular generation to systematically investigate the effects of chemical structure, FR addition amount, curing agent type, and curing agent ratio on the limiting oxygen index (LOI) of EP composites. The optimized predictive model achieved a coefficient of determination (R²) of 0.83, identifying critical structure–property relationships, including molecular weight, heavy atom count, aromaticity, epoxy equivalent, curing dynamics, and crosslinking parameters. From a library of > 860,000 candidate molecules, 8 high-potential FRs − panning monomeric and compounded systems − were computationally prioritized. Experimental validation demonstrated remarkable performance metrics: a monomeric FR achieved an LOI of 36.4 %, while a compounded system delivered a record LOI of 42.5 %, positioning it within the top 0.4 % of reported FR benchmarks. Crucially, this breakthrough was achieved with unprecedented cost-efficiency − a 95.6 % reduction in per-kilogram FR material cost and 30 % formulation cost reduction for EP composites. The ML framework exhibited robust predictive accuracy, achieving a Mean Absolute Error (MAE) of 1.69 % for monomeric FR systems in LOI prediction, while demonstrated 4.81 % for compounded FR systems. This ML framework not only bridges critical gaps in flame retardancy science but also establishes a transformative paradigm for intelligent architectural material design with unprecedented cost-efficiency.