Abstract
<title>Abstract</title> <p>Predicting crystal structures from chemical formulas remains a bottleneck for downstream simulations requiring explicit atomic coordinates. We present an automated template-based composition-to-structure workflow combining periodic-descriptor-based template retrieval, a newly developed stoichiometry-aware substitution engine, and MatterSim relaxation. The descriptor is used both to rank compositionally similar templates and to predict space-group probabilities as a structural prior. We benchmark the workflow on Materials Project and further evaluate it using the experimental AtomWork-Adv. database. Under unconstrained retrieval, the space-group match rate on the valid Materials Project subset is 40.1%; using the top-ranked predicted space group increases it to 63.0% (+ 22.9 percentage points) with comparable substitution success. The same trend is reproduced on AtomWork-Adv. Broader top-K space-group sets improve template coverage but increase competition from descriptor-similar templates. The workflow provides an interpretable, traceable, and reproducible route from chemical formulas to candidate inorganic crystal structures.</p>