Bayesian Optimization of flame-retardant performance in a high-Tg epoxy resin system

Polymeric materials are widely used due to their mechanical properties and cost-effectiveness, but their inherent flammability requires effective flame-retardant additives to meet safety standards. Optimizing multi-component flame-retardant formulations is challenging due to the vast experimental space. This study applies Bayesian Optimization (BO) to optimize flame-retardant formulations in high glass transition temperature (Tg) epoxy resins. Aluminum diethyl phosphinate (AlPi) was systematically combined with three synergists: zinc stannate (ZnSt), a silicone-based additive (DowSil), and low-melting glass frits (Ceepree). BO-guided experimental design expanded from 16 initial formulations to a total of 28, minimizing the Maximum Average Rate of Heat Emission (MARHE) under the constraint of Total Smoke Production (TSP) < 17 m 2 using the epsilon-constraint method. BO revealed non-linear synergistic interactions: ZnSt significantly reduced smoke production while AlPi effectively lowered heat release. The optimized formulation (BO7) achieved the lowest MARHE (122 kW/m 2 ) while maintaining acceptable smoke levels, establishing a new Pareto front. The results demonstrate the effectiveness of BO in accelerating the development of synergistic, halogen-free flame-retardant polymer systems, offering a scalable and sustainable approach to polymer formulation design. • Workflow of Bayesian Optimization applied to flame-retardant formulation design in high-Tg epoxy resins. • Starting from initial experimental data, BO iteratively suggests new formulations,which are processed, characterized, and analyzed for heat release (MARHE) and smoke production (TSP). • The optimization cycle refines the dataset, revealing synergistic effects and guiding the discovery of optimal halogen-free formulations.