Accurate evaluation of Li-ion battery (LiB) safety conditions can reduce unexpected cell failures, facilitate battery deployment, and promote low-carbon economies. Despite the recent progress in artificial intelligence, anomaly detection methods are not customized for or validated in realistic battery settings due to the complex failure mechanisms and the lack of real-world testing frameworks with large-scale datasets. Here, we develop a realistic deep-learning framework for electric vehicle (EV) LiB anomaly detection. It features a dynamical autoencoder tailored for dynamical systems and configured by social and financial factors. We test our detection algorithm on released datasets comprising over 690,000 LiB charging snippets from 347 EVs. Our model overcomes the limitations of state-of-the-art fault detection models, including deep learning ones. Moreover, it reduces the expected direct EV battery fault and inspection costs. Our work highlights the potential of deep learning in improving LiB safety and the significance of social and financial information in designing deep learning models.

Reaction rates at spatially heterogeneous, unstable interfaces are notoriously difficult to quantify, yet are essential in engineering many chemical systems, such as batteries¹ and electrocatalysts². Experimental characterizations of such materials by operando microscopy produce rich image datasets^3–6, but data-driven methods to learn physics from these images are still lacking because of the complex coupling of reaction kinetics, surface chemistry and phase separation⁷. Here we show that heterogeneous reaction kinetics can be learned from in situ scanning transmission X-ray microscopy (STXM) images of carbon-coated lithium iron phosphate (LFP) nanoparticles. Combining a large dataset of STXM images with a thermodynamically consistent electrochemical phase-field model, partial differential equation (PDE)-constrained optimization and uncertainty quantification, we extract the free-energy landscape and reaction kinetics and verify their consistency with theoretical models. We also simultaneously learn the spatial heterogeneity of the reaction rate, which closely matches the carbon-coating thickness profiles obtained through Auger electron microscopy (AEM). Across 180,000 image pixels, the mean discrepancy with the learned model is remarkably small (<7%) and comparable with experimental noise. Our results open the possibility of learning nonequilibrium material properties beyond the reach of traditional experimental methods and offer a new non-destructive technique for characterizing and optimizing heterogeneous reactive surfaces.