Mapping artificial intelligence for science with open metadata reveals shifting evidence regimes across scientific domains

Artificial intelligence for science (AI4S) is often discussed as a unified transformation, although scientific domains differ markedly in how model outputs are tested, validated, and justified. We map an audited corpus of N=2,784 records (2014–2025) retrieved from Crossref and screened with a frozen, versioned rulebook, and compare evidence regimes in protein science, materials design, drug discovery, closed-loop experimentation, and mathematical reasoning. Using term co-occurrence, reference co-citation, burst detection, growth models, and collaboration networks, we identify three robust patterns: (i) rapid but decelerating growth (Poisson rate 1.381, 95% CI 1.361–1.402, declining from 1.537 in 2014–2019 to 1.320 in 2019–2025); (ii) a coherent term-co-occurrence structure (75 nodes, modularity Q=0.404 ) organized around four measured streams—representation learning, generative and inverse design, closed-loop experimentation, and neuro-symbolic reasoning—with the generative stream acting as the connecting hub; and (iii) a right-boundary rise in large language model terminology (burst state 3) that a year-by-year backtest shows is not a last-observation artefact. The maps are sensitive to metadata availability: abstract coverage is 50.6%, reference coverage 84.3%, and a title-only variant changes the network partition substantially (adjusted Rand index 0.762), so the maps describe the screened open-metadata corpus rather than the total AI4S literature. Cross-domain analysis shows that AI most consequentially shifts scientific bottlenecks from representation and candidate generation toward experimental or formal validation, provenance, and responsibility for claims. We contribute an auditable, open-metadata science-mapping workflow designed for reproducibility and an evidence-regime framework for analysing how epistemic labour is redistributed across models, instruments, formal verifiers, infrastructures, and human experts.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11461912/v1
Primary Topic
scientometrics and bibliometrics research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Mapping artificial intelligence for science with open metadata reveals shifting evidence regimes across scientific domains

Xinliang Wang, Xu Fengchang, Lv Ge, Zhang Na
2 links
Research Square
scientometrics and bibliometrics research
preprint

Mapping artificial intelligence for science with open metadata reveals shifting evidence regimes across scientific domains

Xinliang Wang, Xu Fengchang, Lv Ge, Zhang Na
preprint en

Abstract

Artificial intelligence for science (AI4S) is often discussed as a unified transformation, although scientific domains differ markedly in how model outputs are tested, validated, and justified. We map an audited corpus of N=2,784 records (2014–2025) retrieved from Crossref and screened with a frozen, versioned rulebook, and compare evidence regimes in protein science, materials design, drug discovery, closed-loop experimentation, and mathematical reasoning. Using term co-occurrence, reference co-citation, burst detection, growth models, and collaboration networks, we identify three robust patterns: (i) rapid but decelerating growth (Poisson rate 1.381, 95% CI 1.361–1.402, declining from 1.537 in 2014–2019 to 1.320 in 2019–2025); (ii) a coherent term-co-occurrence structure (75 nodes, modularity Q=0.404 ) organized around four measured streams—representation learning, generative and inverse design, closed-loop experimentation, and neuro-symbolic reasoning—with the generative stream acting as the connecting hub; and (iii) a right-boundary rise in large language model terminology (burst state 3) that a year-by-year backtest shows is not a last-observation artefact. The maps are sensitive to metadata availability: abstract coverage is 50.6%, reference coverage 84.3%, and a title-only variant changes the network partition substantially (adjusted Rand index 0.762), so the maps describe the screened open-metadata corpus rather than the total AI4S literature. Cross-domain analysis shows that AI most consequentially shifts scientific bottlenecks from representation and candidate generation toward experimental or formal validation, provenance, and responsibility for claims. We contribute an auditable, open-metadata science-mapping workflow designed for reproducibility and an evidence-regime framework for analysing how epistemic labour is redistributed across models, instruments, formal verifiers, infrastructures, and human experts.

Research Square
Changzhou Vocational Institute of Light Industry (CN), Shandong Vocational College of Light Industry (CN)
scientometrics and bibliometrics research
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