Combinatorial Precursor Set Prediction for Inorganic Materials Synthesis via Graph-Conditioned Sequence Generation

The realization of novel inorganic materials is fundamentally bottlenecked by the challenge of experimental synthesis planning. Recent data-driven approaches primarily frame precursor selection as a retrieval or ranking problem, relying on historical precedent to propose recipes for new targets. Here, we demonstrate that this precedent-driven paradigm is strictly constrained by a combinatorial ceiling. Evaluating N=20,103 materials reported post-2020 against the pre-2020 literature, we find that while 92.4% of individual precursor ingredients are familiar, the exact required combinations are entirely unseen in 46.8% of cases. Consequently, any extractive or template-based retrieval method faces a hard oracle upper bound of 53.2% exact-match accuracy. To overcome this barrier, we recast synthesis planning as generative reasoning under combinatorial constraints. We introduce a dual-modality sequence-to-sequence generator that dynamically constructs open-ended precursor sets rather than extracting them. By fusing elemental descriptors with 3D geometric embeddings, our approach explicitly resolves polymorph-specific constraints that composition alone underdetermines. On the strict time-split benchmark, our hybrid generative framework achieves a 32.5\% exact-match accuracy overall, more than doubling the performance of state-of-the-art retrieve-and-recombine baselines. Most notably, the model achieves a 21.9\% exact match on genuinely novel precursor combinations, a domain where the theoretical oracle limit for extractive retrieval is strictly zero. By shifting from historically constrained extraction to open-ended generative construction, this work overcomes the fundamental limits of analogy-based synthesis and enables the prospective planning of viable chemical routes for unprecedented inorganic solids.