Pores for Thought: Reconstructing 3D, Multi-Phase Electrode Microstructures with Periodic Boundaries Using Generative Adversarial Networks for Battery Design

The microstructure of porous electrodes significantly impacts the performance of electrochemical energy storage (EES) devices [1,2]. Thus, their morphological optimisation is vital for developing the next generation of EES technologies [3]. Recent improvements in 3D imaging techniques have allowed the characterisation of porous electrodes at a nanoscale [4]. However, a variety of challenges remain, including how to extract the key metrics or “essence” of the microstructure with which synthetic volumes with equivalent properties can be generated. This work implements a machine learning technique called deep-convolutional generative adversarial networks (GANs) to generate 3D synthetic realisations of n-phase electrode microstructures (Figure 1 (a)). The same network architecture is successfully applied to generate two very different three-phase microstructures: a Li-ion cathode and a SOFC anode. GANs implicitly capture the probability distribution function that fully defines the microstructure [5] into a low dimensional space, which is crucial for the follow-up process of optimisation. A statistical comparison between the real and synthetic data is performed in terms of the morphological properties (volume fraction, surface area, triple phase boundary) and transport properties (relative diffusivity), as well as the two-point correlation function. The results show excellent agreement between both datasets and they are also visually indistinguishable. The impact of this work lies in the ability to obtain a fully differentiable function that constitutes a “virtual representation” of the electrode microstructure. We show that with these virtual representations, it is possible to implement physical simulations based on which we can design an optimised porous microstructure. Furthermore, we introduce the possibility of generating spatially periodic microstructures with the same microstructural properties as the real tomographic data (Figure 1 (b)). These periodic microstructures have the potential to significantly reduce the simulated volume required to be considered “representative" and therefore massively reduce the computational cost of the electrochemical simulations necessary to predict the performance of a particular microstructure during optimisation. Keywords : Electrode microstructure, machine learning, generative adversarial networks, optimisation [1] G. Wei, C. Jia, J. Liu and C. Yan, Journal of Power Sources 220, 185 (2012). [2] H. Moussaoui, J. Laurencin, Y. Gavet, G. Delette, M. Hubert, P. Cloetens, T. Le Bihan and J. Debayle, Computational Materials Science. 143, 262 (2018). [3] N. Ni, S.J. Cooper, R. Williams, N. Kemen, D.W. McComb, and S.J. Skinner, ACS Applied Materials and Interfaces. 8(27), 17360 (2016). [4] F.L.E. Usseglio-Viretta, A. Colclasure, A.N. Mistry, K.P.Y. Claver, F. Pouraghajan, D.P. Finegan, T.M.M. Heenan, D. Abraham, P.P. Mukherjee, D. Wheeler, P. Shearing, S.J. Cooper and K. Smith, Journal of The Electrochemical Society. 165, A3403 (2018). [5] A. Gayo-Lombardo, L. Mosser, N.P. Brandon, S.J. Cooper, npj Computational Materials (2020) (Pre-print arXiv https://arxiv.org/abs/2003.11632) Figure 1