EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multimodal Learning with Rashomon Analysis for Battery Discharge Capacity Prediction

Jue Gong, Lejia Wang, Tianqing Zhu, Keith Butler, Sichao Li

Advanced Intelligent Discovery · 15.06.2026

DOI: 10.1002/aidi.70140 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Advanced Intelligent Discovery, online erschienen am 15.06.2026. Verfasst von Jue Gong und 4 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Multimodal learning is increasingly used in materials informatics because electrochemical properties depend on multiscales, e.g., complementary information distributed across composition, local structure, and medium‐range atomic organization. However, multimodal studies are still commonly interpreted through a single selected model, even when several models achieve similar predictive performance. This work addresses that issue through a Rashomon‐aware multimodal framework for discharge‐capacity prediction in a battery‐compound dataset represented by compositional descriptors ( COMP ), crystal‐structure descriptors ( CRY ), and radial‐distribution descriptors ( RDF ). Across the reported benchmarks, the trimodal model achieves the strongest overall fit, outperforming all unimodal and bimodal baselines in , mean squared error (MSE), and root mean squared error (RMSE), while CRY – COMP is the strongest bimodal configuration. Beyond prediction, the study primarily analyzes a set of near‐optimal models, known as a Rashomon set, to examine how explanatory conclusions vary within a narrow performance band. The resulting interpretability analysis shows that explanatory variation is structured rather than arbitrary. Transition‐metal‐related descriptors in COMP, selected bond‐angle‐ and bonding‐related CRY channels, and short‐range RDF shell features recur across near‐optimal models, whereas other signals are more model‐contingent. These results suggest that multimodal fusion improves predictive modeling for heterogeneous battery compounds, and that explanation diversity among near‐optimal models can be treated as a form of structured epistemic uncertainty. This study establishes multimodal learning and Rashomon analysis as a principled pathway from prediction to mechanistic insight in energy materials research.

Zitieren

APA

Gong, J., Wang, L., Zhu, T., Butler, K., & Li, S. (2026). Multimodal Learning with Rashomon Analysis for Battery Discharge Capacity Prediction. Advanced Intelligent Discovery. https://doi.org/10.1002/aidi.70140

Vancouver

Gong J, Wang L, Zhu T, Butler K, Li S. Multimodal Learning with Rashomon Analysis for Battery Discharge Capacity Prediction. Advanced Intelligent Discovery. 2026. doi:10.1002/aidi.70140

BibTeX
@article{Gong2026,
  author = {Gong, Jue and Wang, Lejia and Zhu, Tianqing and Butler, Keith and Li, Sichao},
  title = {Multimodal Learning with Rashomon Analysis for Battery Discharge Capacity Prediction},
  journal = {Advanced Intelligent Discovery},
  year = {2026},
  doi = {10.1002/aidi.70140}
}
Kostenlos starten →