EUVIMEDEUVIMED entdecken
ÜbersichtsarbeitOpen Access · CC BY 4.0

From Prediction to Physical Meaning in Materials AI

Meng Li iD, Pengzhen Lin

ACS Materials Au · 11.11.2026 · 6 (6) · 1209-1216

DOI: 10.1021/acsmaterialsau.6c00170 ↗

Kurz gefasst

Übersichtsarbeit in der Zeitschrift ACS Materials Au, online erschienen am 31.08.2026. Verfasst von Meng Li und 1 weiteren Autor:innen. Beteiligte Einrichtung: Department of Mechanical and Materials Engineering, Queen's University, Kingston, Ontario K7L 2N8, Canada. Die Arbeit nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Canada Excellence Research Chairs, Government of Canada.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Materials artificial intelligence (AI) is increasingly used to support claims that extend beyond property prediction, including physical interpretation, mechanism identification, and design guidance. As the claim shifts from prediction to understanding, the required evidence must shift accordingly. In this Perspective, we argue that physical meaning in materials AI should be treated as a claim-evidence relationship rather than as a property conferred by model architecture or explanation method. Predictive accuracy establishes performance within a tested domain, physical consistency shows compatibility with constraints, and model interpretation reveals how information is used. These are necessary forms of evidence, but they have boundaries, and none by itself justifies a mechanism or design principle. We use a claim-evidence map and validation ladder to identify where evidence boundaries arise and what additional support is required for stronger claims. Examples from out-of-distribution property prediction, text-mined synthesis data, and machine-learned interatomic potentials show how validation domain, data set history, and sampled configuration regime can limit the physical claims a model can support. We then use defect-chemistry-guided materials AI in proton-conducting oxides as a case analysis, where physically defined defect variables allow model outputs to be evaluated across prediction, interpretation, transport mechanism, and domain-limited design descriptors. The central argument is that physical meaning depends on aligning the strength of a materials claim with the evidence available to support it. Progress in materials AI should not be judged by predictive performance alone, but whether models make physical meaning testable.

Schlagwörter

Stichwörter der Autor:innen

claim-evidence relationshipdefect chemistryevidence boundaryexplainable machine learningmaterials artificial intelligencephysical meaningproton-conducting oxides

Autor:innen und Institutionen

  • Department of Mechanical and Materials Engineering, Queen's University, Kingston, Ontario K7L 2N8, Canada.
  • Department of Statistical Sciences, University of Toronto, Toronto, Ontario M5S 1A1, Canada.

ORCID-Profile: Meng Li

Förderung

  • Canada Excellence Research Chairs, Government of Canada

Zitieren

APA

Li, M., & Lin, P. (2026). From Prediction to Physical Meaning in Materials AI. ACS Materials Au, 6(6), 1209-1216. https://doi.org/10.1021/acsmaterialsau.6c00170

Vancouver

Li M, Lin P. From Prediction to Physical Meaning in Materials AI. ACS Materials Au. 2026;6(6):1209-1216. doi:10.1021/acsmaterialsau.6c00170

BibTeX
@article{Li2026,
  author = {Li, Meng and Lin, Pengzhen},
  title = {From Prediction to Physical Meaning in Materials AI},
  journal = {ACS Materials Au},
  year = {2026},
  volume = {6},
  number = {6},
  pages = {1209--1216},
  doi = {10.1021/acsmaterialsau.6c00170}
}
Kostenlos starten →