PRICIN: Principle-Centered Inorganic Retrosynthesis

ICLR 2026 Conference Submission15293 Authors

Keywords: Inorganic retrosynthesis; Retrieval augmentation; Materials science; Lab automation

TL;DR: Chemical rules guided inorganic retrosynthesis.

Abstract: Bridging the gap between what is designable by computational discovery and what is synthesizable in the lab remains a central obstacle for closed-loop materials science. We tackle single-step inorganic retrosynthesis and show that explicit chemical principles are potent inductive biases for learning to plan syntheses. We introduce PRICIN, a principle-centered approach that reformulates precursor planning around two laws: elemental conservation and electron balance. PRICIN embeds stoichiometry and oxidation-state semantics directly into the target representation via two pretraining objectives, including an auxiliary oxidation-state supervision that injects charge awareness. At inference, a lightweight element-wise filter first predicts the required number of precursors and then prunes candidates that violate conservation constraints, yielding explainable, chemically consistent precursor sets without external retrieval or rigid templates. Across the Retrieval-Retro (year-split) and Ceder benchmarks, PRICIN attains state-of-the-art performance on Top-$k$ and combination Top-$k$ metrics, improving over the previous best by +5.59 Top-1 and by up to +19.2 points on Top-$k$. Ablations confirm that oxidation-state supervision and conservation-aware filtering are both necessary and complementary, substantially reducing early-rank errors. The code will be released upon acceptance.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15293