What we know and what we don’t: an audit of experimental solid-electrolyte data, and the patent literature as an untapped source

Machine-learning (ML) models for solid-state electrolytes (SSEs) are trained on experimental ionic-conductivity data, yet public datasets are smaller, more duplicated and noisier than their sizes suggest. We merged the five open Li conductivity datasets we could download (Liverpool Ionics, Laskowski et al., McHaffie et al., OBELiX and the text-mined set of Shon and Min) and audited them in three ways. First, after re-extracting one dataset from its PDF table, normalising compositions and collapsing copies, the four curated datasets shrink from 3,336 nominal entries to 1,481 unique compositions from 540 papers; the CSV rendering most users would download hides a fifth of them. 86% of OBELiX records are exact copies of records in its parent datasets. Second, for 194 compositions reported by two or more papers, the median spread is 3.4-fold, and a report lies on average 0.43 log units from its composition median, nearly three times the spread among eight laboratories measuring identical powder. A systematic offset per paper accounts for 85% of this between-paper variation, and phase labels explain much of the largest spreads, while activation energies vary between papers no more than between laboratories. Third, one gradient-boosting model scores a mean absolute error (MAE) of 0.90 log units under random cross-validation, 1.20–1.42 when entries from the same paper are kept together, 1.89 on a time split and 2.24 when a structure family is held out, worse than predicting the median; five other learners show the same pattern. Concatenating public datasets without deduplication lets copies leak into test folds and lowers the random-split MAE to 0.57, close to the experimental spread. We also built and tested a pipeline that extracts composition–conductivity records with evidence quotes from 100 recent patents in CPC H01M 10/0562. One in five reports measured conductivities, and these documents hold 171 distinct lithium compositions, 92% of them absent from the open datasets; halides make up a third of them against 4% of the literature set, and garnet and perovskite oxides are missing. Against a 25-patent reference set the pipeline recovers about half of the table records at a precision of 0.84; the reference set was drafted by the extraction model and has not yet been checked by hand, so these scores are provisional. We recommend a minimum reporting standard for conductivity, published duplicate maps and versions for datasets, group-aware splits reported against the experimental spread, and release of patent-derived facts with citations only.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-16
DOI
https://doi.org/10.26434/chemrxiv.15012245/v1
Primary Topic
Machine Learning in Materials Science
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

What we know and what we don’t: an audit of experimental solid-electrolyte data, and the patent literature as an untapped source

Henrik Laar, Serhii Vakal
4 links
ChemRxiv
Machine Learning in Materials Science
preprint

What we know and what we don’t: an audit of experimental solid-electrolyte data, and the patent literature as an untapped source

Henrik Laar, Serhii Vakal
preprint en

Abstract

Machine-learning (ML) models for solid-state electrolytes (SSEs) are trained on experimental ionic-conductivity data, yet public datasets are smaller, more duplicated and noisier than their sizes suggest. We merged the five open Li conductivity datasets we could download (Liverpool Ionics, Laskowski et al., McHaffie et al., OBELiX and the text-mined set of Shon and Min) and audited them in three ways. First, after re-extracting one dataset from its PDF table, normalising compositions and collapsing copies, the four curated datasets shrink from 3,336 nominal entries to 1,481 unique compositions from 540 papers; the CSV rendering most users would download hides a fifth of them. 86% of OBELiX records are exact copies of records in its parent datasets. Second, for 194 compositions reported by two or more papers, the median spread is 3.4-fold, and a report lies on average 0.43 log units from its composition median, nearly three times the spread among eight laboratories measuring identical powder. A systematic offset per paper accounts for 85% of this between-paper variation, and phase labels explain much of the largest spreads, while activation energies vary between papers no more than between laboratories. Third, one gradient-boosting model scores a mean absolute error (MAE) of 0.90 log units under random cross-validation, 1.20–1.42 when entries from the same paper are kept together, 1.89 on a time split and 2.24 when a structure family is held out, worse than predicting the median; five other learners show the same pattern. Concatenating public datasets without deduplication lets copies leak into test folds and lowers the random-split MAE to 0.57, close to the experimental spread. We also built and tested a pipeline that extracts composition–conductivity records with evidence quotes from 100 recent patents in CPC H01M 10/0562. One in five reports measured conductivities, and these documents hold 171 distinct lithium compositions, 92% of them absent from the open datasets; halides make up a third of them against 4% of the literature set, and garnet and perovskite oxides are missing. Against a 25-patent reference set the pipeline recovers about half of the table records at a precision of 0.84; the reference set was drafted by the extraction model and has not yet been checked by hand, so these scores are provisional. We recommend a minimum reporting standard for conductivity, published duplicate maps and versions for datasets, group-aware splits reported against the experimental spread, and release of patent-derived facts with citations only.

ChemRxiv
Silicon Labs (United States) (US)
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon