Battery interfaces in the face of machine learning and artificial intelligence

Since its initial commercialization over three decades ago, Li-ion batteries have evolved to become a crucial component of our global energy transformation to reduce carbon emissions, from powering electric vehicles to storing renewable energy in electrical grids. To meet the exploding demands for battery energy storage, there is an urgent need to address the current exploding battery problem. The flammable liquid electrolytes used in conventional batteries pose risks of fires and explosions in the event of a thermal runaway, prompting the development of new batteries that can lower or eliminate such risks. Solid-state batteries hold promise as the next generation of batteries, using solid electrolytes in place of liquid electrolytes, which can bring improvements in not just safety but also battery performance. However, before practical solid-state batteries can be realized, there are many technical challenges that need to be overcome, such as the poor long-term cycling stability of current solid-state batteries. Instabilities arise in the solid-solid interfaces within the battery, especially in the interfaces between the solid electrolyte and the electrodes where decomposition reactions often occur. Experimental efforts to characterize these interfaces face significant obstacles due to the sensitive, buried nature of these interfaces, severely limiting the information obtainable from experiments. Therefore, computational studies of battery interfaces play an important, complementary role in understanding these complex interfaces and designing strategies to mitigate interfacial instabilities. This thesis presents a body of computational work revolving around solid-solid battery interfaces. Included are numerous studies utilizing a spectrum of battery interface modeling techniques, from indirect methods that rely on bulk calculations of single-phases, to direct methods that model the interfaces explicitly, to new machine learning techniques that we developed for interface generation. We studied the phase transformation of the MgxMn2O4 cathode, which was found to self-arrest even in nanoparticles, resulting in phase heterogeneity and detrimental strain fields that lower ion diffusivity. We showed that a weak thermodynamic driving force for phase transformation and unfavorable increases in mismatch strain at grain boundaries can inhibit the nanoparticles from complete phase transformation. We also studied the mechanism of how CF4/Ar plasma treatment increases the interfacial stability of LiCoO2 (LCO) cathodes (with solid electrolytes). We showed that the heavily fluorinated LCO cathode has a strong tendency to form fluoride-rich cathode electrolyte interphases, which have a passivating effect that suppresses further interfacial decomposition, enabling cycling at high voltages. We also studied the role of crystallographic orientation on interfacial stability between the LCO cathode and the Li3YCl6 solid electrolyte. We showed that different LCO facets expose different chemical species at their surfaces, resulting in different interfacial binding strength and stability. Through computational modeling of these interface systems, valuable insights were uncovered to help understand the novel interfacial behaviors observed in experiments, and inform future research efforts. Interfaces can exist in many different configurations that are energetically similar, and certain interface systems can even exhibit significant disorder (unlike simple epitaxial interfaces). Explicit interface modeling requires consideration of the many possible interface configurations, which is very costly in computation resources, making intelligent interface search schemes highly desirable. The interface search problem can be approached using generative modeling, a machine learning technique with the goal of learning the underlying distribution of a dataset such that new data can be sampled from the learned distribution. By improving upon existing generative models intended for simple crystals, we developed a novel generative adversarial network, CryinGAN, that can learn from a dataset of energetically favorable interface structures, to generate new interface structures that are energetically and structurally similar to the training structures. We developed CryinGAN using a LCO-Li3ScCl6 disordered battery interface system, and demonstrated the capability to generate large, complex interface structures with >250 atoms for the first time. Through the combination of generative modeling and machine learning interatomic potentials, which can perform relaxations rapidly with high accuracy, we demonstrated an effective interface search scheme that outperforms random search. Beyond leveraging generative modeling to make advancements in interface generation, we also conversely leveraged interface generation to make advancements in generative modeling. Generative modeling efforts for inorganic materials have mostly focused on simple crystals (usually <=20 atoms), even though the ability to generate larger, more complex materials would massively expand the applicability of generative modeling to a broader spectrum of materials. Also, generative models are usually evaluated based on newly generated, unverified materials using heuristic metrics such as charge neutrality, which provide a narrow evaluation of a model's performance. To address these limitations, we developed a new benchmark for generative models, Dismai-Bench, which uses datasets of disordered interfaces, alloys, and amorphous silicon that have >250 atoms per structure. Dismai-Bench follows the same evaluation framework that was used during CryinGAN's development, where models are evaluated through direct comparisons between training and generated structures. This approach provides meaningful feedback about model performance in learning the complex structural patterns of the disordered material datasets. We showed that generative models performed better with higher expressive power of their material representation. We also demonstrated that, with careful and robust evaluation, a simpler model such as CryinGAN can outperform models with more sophisticated architectures. The Dismai-Bench (and CryinGAN) code and datasets were made openly available to help the development of the next generation of generative models. This thesis makes unique connections between battery interfaces and machine learning, offering novel contributions with broad impacts to both fields. We hope that this thesis will inspire new innovations that find inventive links between battery interfaces, disordered materials, machine learning, generative modeling, and beyond.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Materials Science & Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Yong, Adrian Xiao Bin

Contributors dc:contributor

• Ertekin, Elif
• Perry, Nicola H.
• Schleife, Andre
• Braun, Paul V.

• Computational
• Battery
• Machine learning
• Generative modeling

Statement dc:rights

• Copyright 2025 Adrian Xiao Bin Yong

Language dc:language

en

Handle dc:identifier

https://hdl.handle.net/2142/129653

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/129653