Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 53 of the 53 references it cites were found in a registry or an index; 99 of the 163 sources its claims cite were checked for support.

From Scientific Copilots to Tool-Grounded Autonomy: AI Agents in Simulation-Driven Materials Discovery

Open access
Alviz-Meza, Anibal, Valencia-Arias, Alejandro, Rojas-Flores, Segundo, Diaz, Felix · 2026 · Data

Abstract

Artificial intelligence (AI) agents and large language model (LLM) agents are beginning to move materials discovery beyond isolated prediction tasks and toward tool-grounded workflows that can retrieve prior knowledge, configure simulations, launch calculations, inspect outputs, and decide what to do next. However, adjacent reviews on materials informatics, self-driving laboratories, natural-language processing in materials science, and autonomous chemistry have not isolated simulation-driven materials workflows as a distinct evidence base. This review addresses that gap through PRISMA-guided searches in Scopus (8 May 2026) and Web of Science (15 June 2026) for English-language journal articles published between 2022 and 2026. The combined search returned 232 records; 27 full texts were assessed and 26 studies were included in the final qualitative synthesis after one full-text exclusion. No eligible study was published in 2022 or 2023, indicating that the field emerged only in 2024 and expanded rapidly in 2025–2026. Catalysis and adsorption tasks (n = 6) and alloy design or evaluation (n = 5) dominated the corpus, while specialized multi-agent architectures were the most common pattern (n = 13). Across the included studies, agentic reasoning was most often coupled to workflow orchestration or integration tools, molecular-dynamics or atomistic simulation environments, and materials-data or machine learning screening pipelines; public repositories or archival artifacts were reported in 18 of 26 studies, experimental validation in six, and robotic closed-loop execution in only one study. The strongest evidence came from workflows that grounded language model decisions in simulators, structured databases, or experimentally verifiable outputs rather than in free-form text alone. This review therefore establishes AI agent workflow orchestration as a distinct analytical category within materials discovery and identifies the reporting, validation, and reproducibility conditions required for these systems to function as credible scientific infrastructure rather than as conversational demonstrations.

Overview

CoChat Summary

TL;DR

This PRISMA-guided systematic review synthesizes 26 studies on AI-agent workflows for simulation-driven materials discovery, finding a field that emerged in 2024 and expanded rapidly through 2025–2026. The strongest evidence comes from tool-grounded orchestration—agents coupled to simulators, databases, and analysis pipelines—rather than free-form language reasoning alone.

Key findings

  • The final corpus comprised 26 studies from 232 combined records; no eligible study appeared in 2022 or 2023, with 1 in 2024, 9 in 2025, and 16 in the partial 2026 segment. Catalysis/adsorption (n = 6) and alloy design/evaluation (n = 5) dominated, and specialized multi-agent architectures were the most common pattern (n = 13).
  • Tool grounding was broad but uneven: workflow orchestration/integration appeared in 21 of 26 studies, materials-data or ML screening pipelines in 11, molecular-dynamics or atomistic simulation environments in 10, explicit DFT/electronic-structure tools in 7, porous-material curation in 2, and robotic execution in 1.
  • Quantitative highlights included Adsorb-Agent recovering comparable adsorption energies in 83.7% of benchmark systems and better energies in 35.0% while using only 26.9% of candidate configurations on average; a crystal-generation framework reducing execution-level hallucinations from 100% to as low as 5% and reaching 82% structural match for fully specified prompts; and an alloy-optimization loop improving tensile-strength estimates from 8.26–9.81 GPa to maxima of 13.31–18.70 GPa.
  • Reproducibility and validation remained uneven: public repositories or archival artifacts were present in 18 of 26 studies, experimental validation in 6, high-autonomy workflows in 6, and genuine robotic closed-loop execution in only 1.

Methods

  • Systematic review reported in accordance with PRISMA 2020, with searches in Scopus (8 May 2026) and Web of Science Core Collection (15 June 2026) for English-language journal articles published 2022–2026.
  • Eligibility required an AI agent, LLM agent, autonomous agent, or multi-agent workflow as an operational part of the study; a materials discovery/design/optimization or closely related computational materials task; and a materially relevant simulation, computational modeling, screening, code execution, or domain tool component. Reviews, perspectives, editorials, conference papers, book chapters, and notes were excluded.
  • Screening reduced 232 records to 76 unique English-language articles, then to 27 full texts and 26 included studies after one full-text exclusion. Data extraction was performed by one researcher with the other approving the final corpus; synthesis was comparative and qualitative because the evidence base was methodologically heterogeneous and a pooled meta-analysis was not pursued.

Limitations

  • The paper states that the review was limited to Scopus and Web of Science Core Collection, restricted to English-language journal articles, and constrained by unstable agentic AI terminology; the 2026 segment is partial because searches were executed on 8 May 2026 and 15 June 2026. It also notes that all reported proportions derive from n = 26, so each inclusion or exclusion shifts a proportion by roughly 3.8 percentage points, and that studies indexed only in other repositories or released as preprints may be under-represented.
  • The paper states that the eligibility criteria requiring a simulation, computational, or domain-specific tool component produced a corpus composed solely of tool-based studies, creating a selection effect; the importance of tool grounding should not be interpreted as evidence that it is more prevalent than non-tool-based methods in the general literature.
  • The text makes evident additional constraints: data extraction was performed by only one researcher, the retained studies differ substantially in material domain, toolchain, and evidence maturity, and many papers lacked complete prompt logs, failure traces, scheduler settings, or environment specifications, making fair comparison and distinguishing genuine autonomy from curated demonstrations difficult.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/data11070180

Cites

53 of 53 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

25 Valid · 9 Not valid · 7 Not checked · 89 Own results and interpretations

56 more claims · 13 with their cited sources checked

Create a free account to see all 130 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 50
  • Probably not a substantial claim. 26
  • Borderline: one of the tests for a substantial claim is unclear. 7

Create a free account to see these claims.

Citations

Cites

Show all 41 references

Metadata

Authors
Alviz-Meza, A., Valencia-Arias, A., Rojas-Flores, S., & Diaz, F.
Affiliations
  • Escuela de Posgrado, Universidad Continental, Lima 15302, Peru
  • Universidad de Los Lagos (CL)
  • Universidad César Vallejo (PE)
  • Universidad Autónoma del Perú (PE)
Venue
Data
Publisher
MDPI AG
Published
2026-07-21
Volume
11
Issue
7
Pages
180
ISSN
2306-5729 (ISSN-L)
OpenAlex
W7169861758
Cited by
0
References
41
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Data (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
AI agents, LLM agents, materials discovery, agentic reasoning, workflow orchestration, multi-agent architecture, materials informatics, self-driving laboratories, molecular dynamics simulation, alloy design, catalysis, adsorption, reproducibility, research infrastructure

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.