Synthesizability-Aware Materials Generation with Target Properties via Reinforcement Learning

Generative models have shown remarkable promise in accelerating materials discovery, yet most generated candidates remain synthetically inaccessible, limiting their practical impact. We address this critical gap by fine-tuning pretrained diffusion models through multi-objective reinforcement learning (RL) that explicitly incorporates multiple synthesis-related constraints for the generation of novel crystal structures with targeted properties and experimental synthesizability. A precursor set of commercially available, non-toxic, and low-cost compounds is constructed, and synthesis-planning models are employed to predict precursor availability, with the resulting score incorporated into the RL reward. Synthesis-related filters, encompassing material class, elemental complexity, and synthesizability score, are further integrated to ensure compatibility with solid-state synthesis routes. The results demonstrate that the proposed framework simultaneously optimizes different material properties, including bulk modulus and magnetic density, while satisfying synthesis constraints, effectively steering the generative model toward functionally promising and experimentally accessible crystal structures.