Machine learning in battery management systems: a structured scoping review and classification of 103 peer-reviewed studies, 2019–2026

Abstract

Machine learning (ML) has moved from a promisingaddition to battery research to a load-bearing component ofmodern battery management systems (BMS), yet the validationpractices behind the literature’s accuracy claims have neverbeen systematically measured. This paper presents a structuredscoping review of 103 peer-reviewed journal articles (2019–2026) — assembled from a Crossref-verified pool and classifieduniformly by management function, method family, chemistry,data source, and validation practice — supported by a 14-paper conference sensitivity set. The bibliometric analysis revealsa clear evolution: classical shallow learners dominated 2019–2021, deep sequence models rose sharply in 2022–2023, physics-informed and hybrid model–data approaches became one of thefastest-growing families (24 of 103 papers), and the first large-language-model-based management studies appear in 2025–2026.Health estimation and prognostics remain the largest applicationcluster (35 papers). Three structural weaknesses persist. Only 9of 103 studies demonstrate online or fleet-level operation, while 17rely on within-cell or k-fold splits that overstate generalization;benchmark reliance is narrow, with 23 papers resting on threepublic datasets; and the laboratory-to-field gap remains the field’sdefining challenge. The review’s sharpest finding is that physics-informed architectures, often assumed to improve reliability byconstruction, are validated less rigorously than purely data-driven peers (54% versus 73% strong-protocol share): physicspriors reduce data demand but do not substitute for rigorousout-of-distribution testing. The paper closes with a conference-path sensitivity analysis, a safety taxonomy for foundation-modelcomponents, and a twelve-item reporting checklist for futuremachine-learning BMS studies.