RealBind: a leakage-controlled benchmark for experimental PXRD and Raman structure retrieval
- 1Independent Researcher
Abstract
We present RealBind, a benchmark for retrieving crystal structures from experimental powder X-ray diffraction (PXRD) and Raman spectra. It contains 3,288 curated candidate structures, of which 1,468 have linked measured PXRD patterns, and 16,005 experimental Raman spectra, of which 10,547 are linked to structures. We provide quality-control records, prototype-held-out and composition-held-out splits, and evaluation protocols that state candidate-library size and the chemical information available at retrieval time. On a 492-candidate prototype-held-out library, a contrastive model trained on simulated PXRD reaches 78.5% Recall@10 on simulated queries but 5.3% on measured queries; physics-inspired corruption raises measured-query Recall@10 to 34.4%. A model-free correlation against simulated candidate patterns, with no preprocessing beyond the common Q-grid resampling and the measuredwindow mask, achieves 82.8% Recall@10 on the same test library. In a separate split comparison, 44% of random-split test structures have a structure-matched training counterpart; measured-query Recall@10 is 59.1% on that split versus 20.7% on the prototype-held-out split, although test populations and libraries also differ. Holding the trained model and 420 queries fixed, restricting the candidate library and pre-screening candidates by the ground-truth element set of the query’s structure, used as an oracle chemistry-conditioned diagnostic, raises Recall@10 from 27.1% to 99.1%; 37.6% of filtered queries have a single candidate. A Raman-to-structure baseline reaches 34.7% Recall@10 on heldout mineral spectra against 979 candidates in the pooled evaluation and 45.7% on the 485-candidate test-only library, which includes no validation structures. RealBind provides an experimental test bed for comparing learned methods with simple simulation-based matching under explicit, leakage-aware protocols.
Related
- Hexacoordinate Actinide(IV) Compounds: Benchmark Complexes for Magnetism and Electronic Structure Analysis.
- Language-model-derived training data for polymer glass transition prediction: a controlled benchmark
- Experimental assessment of efficiency degradation and corrosion behavior of centrifugal pumps under laboratory-controlled high-salinity conditions
- Revealing Phase-Controlled Surface Structure in PtNi Nanocatalysts via In-situ Gas Cell Transmission Electron Microscopy
- Structure–function relationships in a square-planar Cu(ii) hybrid halometalate: experimental, DFT, molecular docking, optical, and biological studies
BibTeX
@misc{srivastava2026realbind,
title = {{RealBind: a leakage-controlled benchmark for experimental PXRD and Raman structure retrieval}},
author = {Preyansh Srivastava},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009668/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009668/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - RealBind: a leakage-controlled benchmark for experimental PXRD and Raman structure retrieval
AU - Preyansh Srivastava
PY - 2026
DA - 2026/09/29
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009668/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009668/v1
AB - We present RealBind, a benchmark for retrieving crystal structures from experimental powder X-ray diffraction (PXRD) and Raman spectra. It contains 3,288 curated candidate structures, of which 1,468 have linked measured PXRD patterns, and 16,005 experimental Raman spectra, of which 10,547 are linked to structures. We provide quality-control records, prototype-held-out and composition-held-out splits, and evaluation protocols that state candidate-library size and the chemical information available at retrieval time. On a 492-candidate prototype-held-out library, a contrastive model trained on simulated PXRD reaches 78.5% Recall@10 on simulated queries but 5.3% on measured queries; physics-inspired corruption raises measured-query Recall@10 to 34.4%. A model-free correlation against simulated candidate patterns, with no preprocessing beyond the common Q-grid resampling and the measuredwindow mask, achieves 82.8% Recall@10 on the same test library. In a separate split comparison, 44% of random-split test structures have a structure-matched training counterpart; measured-query Recall@10 is 59.1% on that split versus 20.7% on the prototype-held-out split, although test populations and libraries also differ. Holding the trained model and 420 queries fixed, restricting the candidate library and pre-screening candidates by the ground-truth element set of the query’s structure, used as an oracle chemistry-conditioned diagnostic, raises Recall@10 from 27.1% to 99.1%; 37.6% of filtered queries have a single candidate. A Raman-to-structure baseline reaches 34.7% Recall@10 on heldout mineral spectra against 979 candidates in the pooled evaluation and 45.7% on the 485-candidate test-only library, which includes no validation structures. RealBind provides an experimental test bed for comparing learned methods with simple simulation-based matching under explicit, leakage-aware protocols.
ER -