Porous battery electrodes exhibit complex, multiscale architectures that strongly influence rate capability, degradation pathways, and manufacturability. Yet the mapping from processing parameters to emergent microstructure remains notoriously difficult to predict with physics-based models alone. This talk introduces a suite of generative AI tools developed at Imperial College London that address this challenge by learning microstructural distributions directly from data and enabling rapid exploration of manufacturable design spaces. Building on techniques such as GAN-based dimensionality expansion for synthesising statistically representative 3D volumes from single 2D micrographs, we demonstrate how data-driven generative models can overcome bottlenecks inherent to conventional image-based simulation workflows.

A central part of the talk examines the generation, analysis, and optimisation of microstructures conditioned on realistic processing parameters such as porosity, active-material fraction, and binder behaviour. By embedding these generative models within Bayesian optimisation loops, we show how large numbers of statistically meaningful microstructures can be produced and evaluated efficiently. This enables the identification of high-performance manufacturing conditions without requiring either full 3D imaging campaigns or computationally intensive virtual fabrication pipelines. While generative models can, in principle, hallucinate physically implausible features, we highlight how incorporating tortuosity depth-profiles, phase topology constraints, and microstructural representativeness tests mitigates this risk and yields structures consistent with experimental XCT trends.

The talk then focuses on quantifying tortuosity factor as a function of electrode depth in graded electrodes. Measuring tortuosity as a depth-resolved field offers a unifying perspective that reconciles XCT-derived tortuosity with impedance measurements from symmetric cells containing blocking electrolytes. We show how this depth-dependent tortuosity can parameterise Doyle–Fuller–Newman models more faithfully for graded electrodes, avoiding oversimplified scalar tortuosity assumptions that misrepresent transport limitations in layered architectures.

To close, we present how our spin-out company, Polaron, is deploying these tools to engineers worldwide for microstructural characterisation, optimisation, and design. We also discuss work from the FULL-MAP project on AI agents for autonomous laboratories, including recent results interrogating how large language models encode chemically structured knowledge such as the periodic table, revealing geometric and layered representations that raise both opportunities and cautions for scientific use. 

This combination of generative modelling, rigorous microstructural physics, and autonomous experimentation sketches a path toward rapid, data-driven materials optimisation pipelines for next-generation electrochemical technologies.