Spec2Prop-InorgBench: An Open Raman--XRD Benchmark and AI-Assisted Screening Workflow for Inorganic Materials

<title>Abstract</title> <p>Open Raman and X-ray diffraction (XRD) resources are valuable for inorganic-material characterization, but differences in file formats, spectral quality, and metadata make reproducible machine-learning evaluation difficult. We present Spec2Prop-InorgBench, an open benchmark that links raw spectra with inorganic material families and database-derived property labels. The benchmark contains 4,118 usable Raman spectra, 4,027 clean inorganic samples, 1,844 formula-matched property records, and 1,462 Raman--XRD linked samples. A runtime preprocessing pipeline converts each spectrum to a standardized 2,048-point representation. Family classification combines PCA features, chemistry-aware spectral descriptors, and train-only prototype-similarity features using a LightGBM-DART classifier, evaluated on an RRUFF-identifier-grouped 70/15/15 hold-out split. The nine-family screening model achieved 66.06\% accuracy and 59.21\% macro-F1 on the held-out test set, with 87.42\% top-3 accuracy. Confidence thresholding at 0.8 increased accepted-sample accuracy to 89.23\% at 43.05\% coverage. For property prediction, descriptor-only XGBoost achieved macro-F1 scores of 75.46\% for band-gap class, 83.18\% for metal/non-metal classification, and 90.98\% for formation-energy class. Spec2Prop-InorgBench provides reproducible benchmark tasks together with the confidence-aware Spec2Prop-Screen workflow for rapid candidate prioritization, intended to assist materials screening rather than replace phase-resolved identification or direct experimental property measurements.</p>