2025 AIChE Annual Meeting

(661d) Enhancing Quantitative Analysis of Lithium-Ion Battery Materials Microscopy through Machine Learning Image Processing

Authors

Samuel Degnan-Morgenstern - Presenter, University of Connecticut
Alexander Cohen, The University of Chicago
Advanced modern imaging techniques have enabled deep insight into the microscopic and nanoscale dynamics of ion transport and reaction kinetics within lithium-ion batteries. In particular, Lim et al. [1] applied scanning transmission X-ray microscopy (STXM) to visualize local lithium composition and characterize spatial heterogeneity within primary lithium iron phosphate particles. Zhao et al. [2] successfully coupled this image dataset to a model-based framework to learn the governing physics through phase field modeling and PDE-constrained optimization. In these studies, experimental noise during the image capture process introduces notable statistical uncertainty into the measured lithium composition and the learned physical model. In this work, we demonstrate the potential of recent deep-learning based image denoising algorithms to enhance data quality and reduce uncertainty within model-based image processing pipelines.

We first verify the viability of machine learning denoising algorithms to improve PDE-constrained optimization on noisy image datasets. A synthetic dataset is generated by simulating the Cahn-Hilliard equation and applying noise sampled from Gaussian, Poisson, and Impulse (Salt-and-Pepper) distributions at increasing noise intensity levels. We apply unsupervised denoising algorithms to the synthetic data and ensure qualitative agreement with the ground truth data. We successfully apply PDE-constrained optimization to extract the underlying governing equations with significantly less uncertainty on the denoised data relative to the noisy images. Additionally, a systematic framework for deploying image denoising algorithms on real X-ray microscopy data is developed by creating heuristics for dataset curation and refining existing image processing pipelines. We demonstrate a reduction in structured noise within the imaged primary particles, revealing hidden heterogeneous domains. Our results show strong qualitative agreement with high-resolution X-ray ptychography. This work establishes the capacity of modern deep learning denoising techniques to improve quantitative analysis of microscopy experiments within material science and engineering.

References:

[1] J. Lim, Y. Li, D. H. Alsem, H. So, S. C. Lee, P. Bai, D. A. Cogswell, X. Liu, N. Jin, Y.-S. Yu, N. J. Salmon, D. A. Shapiro, M. Z. Bazant, T. Tyliszczak, and W. C. Chueh. Origin and hysteresis of lithium compositional spatiodynamics within battery primary particles. Science, 353(6299):566–571, August 2016.

[2] H. Zhao, H. Dean Deng, A. E. Cohen, J. Lim, Y. Li, D. Fraggedakis, B. Jiang, B. D. Storey, W. C. Chueh, R. D. Braatz, and M. Z. Bazant. Learning heterogeneous reaction kinetics from X-ray videos pixel by pixel. Nature, 621(7978):289–294, September 2023.