From Task-Specific AI to Battery Foundation Models

Abstract

Artificial intelligence (AI) is influencing battery research and life-cycle management, spanning materials discovery, electrode and cell design, manufacturing optimization, diagnostics, safety monitoring, recycling, and certification. These advances are enabled by growing battery datasets and learning methods that extract patterns from complex electrochemical systems. Yet most current models remain task-specific and are trained on fragmented datasets, limiting transferability across chemistries, cell formats, operating conditions, and life-cycle stages. Battery foundation models (BFMs) offer a promising but emerging paradigm for battery intelligence. They may be formulated as large, pre-trained, and adaptable architectures that integrate multimodal battery data with physics-informed priors to represent electrochemical structures, states, and dynamics. Unlike direct analogues of general large language models, BFMs should be understood as physics-aware foundation models designed for scientific reasoning under multiscale physical constraints, heterogeneous and incomplete data, and stringent safety and reliability requirements. By combining self-supervised learning with mechanistic electrochemical models, BFMs may learn more interpretable and transferable representations across materials, formats, and duty cycles. If sufficiently validated, such representations could support materials discovery, manufacturing, diagnostics, recycling, and certification. This Review discusses challenges in data fragmentation, validation, interpretability, and computational sustainability, and outlines a roadmap toward responsible BFMs and integrated battery-domain ecosystems across the lifecycle.