Optimizing battery technologies often requires a significant amount of time-consuming and costly trial and error. Artificial intelligence (AI) and machine learning (ML) could accelerate the optimization of battery technologies. The design of useful predictive AI/ML approaches needs proper training using high-quality datasets. Could the data already available in the battery-related scientific literature—over 28,000 articles about lithium-ion batteries, for example—be used to train such AI/ML models?
Alejandro A. Franco, Université de Picardie Jules Verne, Amiens, France, and colleagues have investigated to which extent the battery-related data available in scientific publications is complete. The team developed a text-mining algorithm that looked for the presence of certain keywords in a set of 13,000 lithium-ion and sodium-ion battery-related scientific articles spanning from 1990 to 2019. The researchers found that a majority of the scientific articles do not report important parameters that could help peers to reproduce the reported experimental data. For example, ca. 85 % of the articles do not report electrode mass loading, ca. 97 % do not report porosity, and ca. 96 % do not report the electrolyte volume.
These results indicate that the battery community might be facing a reproducibility crisis, and that open data and standardization actions are urgently needed to improve the quality and completeness of the datasets reported in the scientific literature. This could make battery research more efficient and allow the emergence of AI tools that can automatically navigate and exploit the reported data for accelerated battery optimization.
- What Can Text Mining Tell Us About Lithium‐Ion Battery Researchers’ Habits?,
Hassna El‐Bousiydy, Teo Lombardo, Emiliano N. Primo, Marc Duquesnoy, Mathieu Morcrette, Patrik Johansson, Patrice Simon, Alexis Grimaud, Alejandro A. Franco,
Batteries Supercaps 2021.