Source-linked AI summary
LitCurate: A Configuration-Driven AI-Assisted Framework for Scientific Database Construction with an Application to Lower-Mantle Equation-of-State Data
Abin Shakya, Wilson Samuels, Dominica Wilson, Gioia A. Marchi, Israa Draz, Chenxing Luo, Renata M. Wentzcovitch
TL;DR
Scientific literature contains valuable results that remain difficult to discover, interpret, and structure for reuse. LitCurate provides an open-source, configuration-driven workflow combining discovery, screening, full-text processing, and schema-guided LLM extraction with retained provenance. Applied to lower-mantle EOS studies, it produced 1,334 entries from 205 papers in a searchable database.
Problem
Scientific results remain locked in publications, while relevant studies must be discovered and reported quantities extracted with enough context for reuse.
Method
LitCurate uses an auditable, stage-wise, configuration-driven workflow that integrates literature discovery, screening, full-text preparation, schema-guided extraction, and provenance retention.
Results
1,334 EOS entries from 205 papers link reported parameters to phases, compositions, formulations, methods, constraints, and provenance labels.
Takeaways & Limitations
LitCurate provides a reusable approach for transforming literature into traceable, machine-readable resources for scientific analysis and computational modeling.
Takeaways & Limitations
The workflow's search terms, screening logic, and extraction targets remain tailored to particular research questions, while relevant papers can be absent even when text extraction is accurate.
Abstract
from arXiv · showhide
The growing scientific literature contains decades of experimental and computational results that could support data-driven and physics-based modeling, yet much of this infor- mation remains locked in publications and is not readily usable for large-scale analysis or sci- entific software. Building structured databases from the literature is particularly challenging whenrelevantstudiesmustfirstbediscoveredamonglargecollectionsofpapersandreported quantities must be extracted with enough scientific context to remain usable. We present LitCurate, an open-source framework for building scientific databases from the literature using large language models within an auditable, stage-wise curation workflow. LitCurate integratesliteraturediscovery, relevancescreening, full-textprocessing, andstructuredinfor- mation extraction while retaining intermediate results and provenance, allowing researchers to inspect and revise individual stages rather than treating automated curation as a black- box process. We apply LitCurate to construct an equation-of-state database of lower-mantle and lower-mantle-relevant high-pressure mineral phases from experimental and theoretical studies, comprising 1,334 entries from 205 papers. The resulting dataset links reported equation-of-state parameters to mineral phases, compositions, equation formulations, meth- ods, and parameter constraints, and labels values as source-reported or citation-reported when provenance can be determined. The records are available through a searchable web application. By connecting scientific literature to traceable, machine-readable data, LitCu- rate provides a reusable approach for transforming accumulated literature into resources for scientific analysis and computational modeling.
1 Introduction
Scientific results remain difficult to reuse because relevant papers must be discovered and reported quantities interpreted with scientific context, heterogeneous conventions, and provenance. LitCurate addresses this gap with an open-source, configurable, auditable workflow that connects literature discovery to structured EOS data extraction.
- Decades of experimental and computational results remain locked in publications rather than readily usable for large-scale analysis or scientific software.
- EOS parameters for lower-mantle minerals vary across formulations, units, reference conditions, and fixed versus fitted parameters.Reported quantities include V0, K0, and K′0 from experimental compression studies and first-principles calculations.
- Building literature-derived databases remains labor-intensive, difficult to maintain, and hard to reproduce when extraction decisions are not recorded.
- Reusable records must retain phase, composition, method, constraint metadata, and literature provenance alongside extracted values.These fields support comparison across studies and transparent downstream model updates.
- Existing extraction platforms often assume a preassembled document collection or provide modular extraction rather than complete discovery-to-export curation.
- Scientific figures remain difficult to extract reliably, often requiring substantial manual annotation and lacking reliability for database construction.
- LitCurate combines literature discovery, screening, full-text preparation, schema-guided extraction, and export in an auditable, stage-wise workflow.Intermediate artifacts and a run ledger support inspection, revision, and resumption, while configuration files hold domain knowledge.
- The application produced 1,334 entries from 205 papers linking EOS parameters to phase, composition, formulation, method, constraints, and provenance labels.Records are available through a searchable web application.
2 Methodology
LitCurate implements literature-to-database curation as configurable stages that begin with a research goal and proceed through discovery, screening, acquisition, conversion, extraction, and export. The workflow preserves intermediate artifacts, validation metadata, and provenance so runs can be inspected, resumed, and traced.
- LitCurate decomposes literature-to-database curation into a fixed sequence of stages that can run individually or as a full pipeline.Interrupted runs resume from the last completed stage.
- A frozen per-run configuration preserves the settings used for completed runs and supports targeted downstream reprocessing after criteria or prompts change.
- Users define research goals, screening criteria, extraction schemas, and prompts in configuration files rather than embedding domain knowledge in pipeline code.The configuration-driven design supports adaptation across disciplines.
- Query planning expands a natural-language research goal into focused searches, whose OpenAlex results are preserved, deduplicated, filtered, and ranked.Ranking combines search frequency, relevance, and citation information.
- Abstract screening records keep-or-exclude decisions with justifications before PDF acquisition and conversion to extraction-ready markdown.PDFs were obtained for 918 of the 1,030 papers retained after abstract screening.
- Two post-conversion gates use regular-expression signals and full-text screening to avoid unnecessary LLM extraction and assess complete documents.
- Schema–prompt pairs instruct the LLM to return JSON records that are validated against declared fields, types, and required properties.Each result is stored with schema version, model identifier, timestamp, validation verdict, and errors.
- Per-paper extraction envelopes are merged into one database while retaining links through manifests to individual extractions, source PDFs, and original queries.
3 Results
The curation run produced a 1,334-entry EOS database from 205 papers and evaluated extraction against 220 manually curated records from 50 papers. Commercial models generally provided stronger record coverage, while field-level extraction was strongest for numeric parameters and weakest for EOS-model identification.
- Curation run: 1,334 EOS entries from 205 papers were compiled after search, screening, PDF acquisition, conversion, and structured extraction.An additional 95 papers produced no exportable records, and 12 papers failed during extraction.
- Evaluation design: 220 manually curated EOS records from 50 papers formed the ground truth for evaluating extraction quality.The records were labeled for phase, composition, EOS model, V0, K0, and K′.
- Model evaluation: Claude Fable 5 recovered 89.1% of ground-truth records and achieved 95.7% accuracy, 99.4% recall, and 97.6% F1 at field level.Its unmatched-record rate was 10.0%.
- Field-level results: Numeric V0, K0, and K′ fields had high F1 and near-perfect recall, whereas the EOS-model field was weakest across evaluated models.The paper links the weaker EOS-model performance to formulations often being applied without being named alongside reported values.
- Curated database: The database contains 679 experimental entries (51%) and 342 computational entries (26%), with bridgmanite as the largest phase group at 348 entries (26%).Periclase contributed 221 entries (16%) and davemaoite 181 entries (13%).
- Scientific plausibility: Extracted parameter distributions clustered around consensus values, including median K0 values of 256 and 255 GPa for experimental and computational bridgmanite.Periclase had a median K0 of 160 GPa for both classes, and fitted K′ had a median of 4.1 with 93% between the reported bounds.
4 Discussion
LitCurate is a reusable, configuration-driven package that records staged literature curation decisions and exports schema-validated databases. Its lower-mantle application shows that screening can reduce expensive extraction workloads, but retrieval and figure-only content remain coverage boundaries.
- Framework design: LitCurate defines research goals, screening criteria, schemas, and prompts in configuration rather than modifying core pipeline stages.This supports adapting the package to new literature-derived database tasks.
- Framework design: The workflow screens papers at two text granularities, records keep-or-drop decisions, and exports a schema-validated database with per-record provenance.Intermediate results are retained so individual stages can be inspected or revised.
- Curation efficiency: Of up to 12,000 OpenAlex query-result records, most did not reach extraction because merge-and-rank, abstract screening, a signal gate, and full-text screening reduced the candidate set.This funnel concentrates expensive LLM extraction on the most relevant documents.
- Coverage boundary: On a 50-paper open-access benchmark, 37 studies appeared in the exported database, and most remaining omissions occurred during retrieval rather than extraction.A relevant paper can therefore be absent even when extraction from converted text is accurate.
- Model deployment: Commercial models were reserved for papers surviving screening, while a locally served open-weight model handled high-volume screening without API cost.The evaluated commercial models performed the production extraction stage well enough for use, with trade-offs among coverage, field-level performance, and cost.
- Database utility: The database retains phase, composition, method, formulation, unit, fitting-constraint, and source-versus-citation provenance metadata.These fields support isolating comparable subsets, restricting analyses to source-reported records, and tracing values to source text.
- Limitations: Extraction reads converted text and tables, so values appearing only as plotted symbols or unlabeled curves are missed.Automated figure extraction is described as not yet mature enough for scientific database construction.
- Limitations: Automatic PDF retrieval is limited to open-access locations returned by Unpaywall and OpenAlex, and open-access flags do not always resolve to downloadable files.Missed copies can be added manually.
5 Conclusions
LitCurate provides a reusable, configuration-driven workflow for converting literature into traceable, schema-validated scientific databases. Its lower-mantle EOS application produced a substantial dataset with quantified retrieval and extraction performance, while figure-based extraction remains outside the current workflow.
- Framework: LitCurate transforms research objectives into structured, schema-validated literature databases through transparent, inspectable stages.The workflow records keep-or-drop decisions and preserves intermediate outputs and provenance.
- Framework: Configuration files define research goals, screening criteria, extraction schemas, and prompts, allowing adaptation across scientific questions without modifying the core pipeline.The framework also supports multiple LLM providers and resumable, auditable curation runs.
- EOS application: 1,334 entries from 205 papers link EOS parameters to phases, compositions, methods, formulations, fitting constraints, and provenance labels.Values may be marked source-reported, citation-reported, or of unknown provenance.
- Evaluation: 74% benchmark coverage was achieved, with 37 of 50 relevant open-access papers represented and most omissions occurring during literature retrieval.The evaluation used 50 papers known to contain relevant EOS parameters.
- Scope and future work: The current workflow excludes quantitative extraction from complex scientific figures because models are not yet sufficiently accurate for that task.Future extensions could add broader property schemas, uncertainty estimates, expert validation, and figure extraction.
Funding
The work was supported primarily by the Gordon and Betty Moore Foundation, with additional support for two authors from the U.S. National Science Foundation.
- Funding: The Gordon and Betty Moore Foundation primarily supported this work through Award GBMF12801.The award DOI is 10.37807/GBMF12801.
- Funding: RMW and WS also received support from the U.S. National Science Foundation through Award EAR-2506448.
Code availability
LitCurate is distributed as a Python package with documented software, hardware, and LLM requirements, and its source code and documentation are publicly available.
- Software: LitCurate is identified as a code/library implemented in Python 3.10 or later.
- Hardware: A CPU is required, while a CUDA-capable GPU is recommended for large-scale PDF conversion.
- Software: The package requires listed Python dependencies and an Anthropic- or OpenAI-compatible LLM API, with optional local serving through Ollama or vLLM.
- Access: Source code is available under the MIT license, and documentation is provided at litcurate.com.
Data availability
The lower-mantle EOS database contains 1,334 entries from 205 papers and is deposited at Zenodo, with column definitions documented separately.
- Database: 1,334 database entries were compiled from 205 papers for the lower-mantle equation-of-state application.
- Access: The dataset is available through Zenodo at DOI 10.5281/zenodo.22118629.
- Documentation: Column definitions for the deposited data are provided in Table S1.
Supplementary material
The supplementary figures detail how LitCurate discovers candidate literature and prepares papers for screening and extraction. Discovery proceeds from query generation through deduplication and ranking, while preparation includes PDF conversion and signal-based filtering.
- Literature discovery expands the research goal into search queries, executes them across publication-year slices, and merges, quality-gates, and ranks overlapping results.
- Screening and preparation acquire PDFs, convert them to cleaned markdown, apply a regular-expression gate for EOS-parameter signals, and screen remaining papers at full text.
S1. Configuration, schema, and extraction prompt
The study configures LitCurate to retrieve and extract static equation-of-state parameters for lower-mantle minerals. Its search and screening prompts favor recall, while schema–prompt pairs preserve reported values, units, provenance, and determination status.
- Configuration and objective: The configuration targets static EOS parameters for lower-mantle phases, especially V0, K0, and K0’.Target phases include bridgmanite, ferropericlase, post-perovskite, CaSiO3 perovskite, and related deep-mantle phases.
- Configuration and objective: Literature discovery uses 20 queries against OpenAlex, with year slices, union-deduplication, ranking, and a top-k candidate limit.The configuration sets top_k to 2000 and allows one query hit to qualify a candidate.
- Screening prompt: Abstract screening keeps papers matching EOS, compression, bulk-modulus, or high-pressure mineral signals even when numeric values are absent from abstracts.The prompt also retains papers that may tabulate previously published parameters and favors recall over precision.
- Screening prompt: Full-text screening retains numeric or clearly tabulated V0, K0, and/or K0’ values, including new fits, cited parameters, and fixed or assumed K0’.Papers are excluded only when no plausible extractable static EOS content exists, and uncertain cases are kept.
- Schema and extraction: The extraction schema stores reported EOS values and units alongside phase, composition, structure, sample, model, method, reference conditions, provenance, evidence, and confidence.Values remain verbatim from the paper, with canonicalization deferred downstream.
S2. Zenodo xlsx deposit
The Zenodo supplementary deposit defines the spreadsheet columns and records the scale and representation of the lower-mantle static EOS database. Reported values are preserved from papers, while selected standardized values are converted after export.
- 1,334 entries from 205 papers spanning 1990–2025 are deposited in the Zenodo Excel file.
- Columns named Reported* store values extracted from the paper.
- V0(cm3/mol) and K0(GPa) are converted after export using only the declared unit.
- Empty phase names are recorded as Unspecified.