AI for Energy Storage

The development of next-generation batteries is bottlenecked by slow, manual scientific workflows. This presentation explores how AI can fundamentally transform how we characterize, design, and optimize battery materials. Drawing on real industrial case studies, we’ll demonstrate how Polaron’s platform combines automated characterization, AI reconstruction, and intelligent design tools to deliver faster insights and higher-performing products—closing the loop between idea, lab, and factory.

Artificial intelligence and machine learning have transformed battery research, progressing integration of computational techniques with chemistry/physics. Starting with quantitative models, progressing through multiphysics models and "digital twins", we now see AI/ML being used as primary research tools. Without physics-based models and real-world test data, we incur the risk that "models training models" loses fidelity. This presentation focuses on the integration of AI/ML, physics, and testing to reduce overtraining risk.

The manufacturing process of batteries has 140 steps and almost 600 variables, which make the optimisation of battery performance against the manufacturing process very challenging. AI techniques provide an opportunity to formulate and understand the impact of key variables. Such AI-based practices are used to make the product performance predictable and reduce the number of tests and experiments needed for its design and optimisation.

Manufacturing anomalies in battery cells can compromise safety and longevity of energy storage systems. Scanning acoustic microscopy offers a promising approach for rapid, cost-effective, and reliable in-line inspection of battery cells. Unsupervised feature extraction from ultrasonic time-series data, combined with clustering algorithms, enables automated anomaly identification, localisation, and classification, reducing production waste and enhancing battery safety.

Industrial battery-sorting processes require fast, non-destructive, and cost-effective diagnostics, but conventional methods are often too slow, intrusive, or resource-intensive. To address this, CIDETEC explored simple voltage measurements as rapid, non-destructive signals for evaluating unknown lithium-ion cells. Explainable machine-learning strategies extracted informative features for reliable chemistry classification. Additionally, voltage data collected at different C-rates supported state-of-health (SOH) estimation and revealed degradation trends relevant to cell reuse. By combining lightweight measurements with data-driven analysis, this methodology offers a scalable, practical solution for automated, high-throughput battery diagnostics, minimising reliance on historical data or extensive laboratory testing.

Field data from batteries offers an enormous and often underused opportunity to accelerate innovation in battery technology. In this talk, I will explore the key challenges and emerging opportunities associated with leveraging real-world operational data throughout the battery development cycle. I will discuss what battery field data can reveal about performance, aging mechanisms, reliability, and user-specific behavior, and how these insights differ from traditional laboratory testing. I will also highlight the analytical and AI-based tools that allow us to process large and noisy datasets, build predictive models, and extract actionable patterns.

Phase transitions in battery electrodes govern rate performance, safety, and lifetime, yet their underlying thermodynamics are typically accessible only through specialised cells and costly operando characterisation. Here, we introduce Bayesian model-integrated neural networks (BMINN) to reconstruct thermodynamically consistent electrode chemical potentials and Gibbs free energies directly from cycling data. Using BMINN, we recover detailed free-energy landscapes that accurately capture staging structures, energy barriers, and phase-separation dynamics. Applied to graphite electrodes, the reconstructed thermodynamics reveal transient phases and heterogeneous intercalation pathways that are otherwise experimentally elusive and remain inaccessible to existing models based on open-circuit potential fits or porous electrode theory.

Develop a rapid, generalised diagnostic that quantifies degradation modes of batteries aged under various conditions as well as  tp remove extensive aging experiments for AI training and employ the identified modes to advance derating control as well as highlight the value of AI in identifying degradation modes and mitigating battery degradation.

Fast Charge management is a must for Electrical Vehicles (EV) to democratize their usage. Indeed, charging time is the companion criteria with embedded energy to increase EV acceptance. Discussion regarding experimental  “high throughput screening” like approach is presented with its limitation. Then Fast charge digital twin approach is presented with use of ROM-AI to connect EV fast charge time to cell design and material choice. Finally, AI capability for design exploration and guidance towards global design optimization.

We highlight the role of differential curves as vital signs of battery health. Moving beyond low-rate characterisation, we show that data-driven methods can uncover patterns from high-rate differential capacity. Explainable AI attributes poor cell health to phase transitions in active materials, linking material behavior to operational outcomes. Using AI to connect data patterns to electrochemical principles supports the diagnosis of battery health and guides design choices for longer battery lifetime.

Li-metal batteries offer exceptional energy density but face safety and cycle life challenges. GBatteries’ Intelligent Battery Management System uses adaptive pulse-based control to enhance performance, reduce degradation, and enable prediction, detection, and prevention of safety events. Validated in drone and aerospace applications, it improves runtime by up to 63%. This session explores how intelligent control accelerates the safe adoption of Li-metal batteries for electric mobility.

This work unifies physics-based electrochemo-mechanical modelling with machine-learning tools to rapidly identify battery safety risks, including defective cells, internal short circuits, and thermal-runaway precursors. Large datasets generated from coupled mechanical and electrochemical simulations enable accurate risk classification across states of charge, loading conditions, and cell formats. The work demonstrates a powerful physics-informed, data-driven approach for real-time battery safety assessment.

Many issues in design, monitoring, and optimisation of battery storage systems are based on time series, for example, from sensor signals. The real systems behind the sensor are inherently inertial. In battery cells, these inertia effects often occur simultaneously on different scales (e.g. electrochemical processes and thermal inertia). AI approaches must therefore be able to deal with inertia effectively. Possible solutions and applications will be presented in the lecture.