Data-driven understanding and forecasting of electrochemical systems

Electrochemical energy systems such as batteries and supercapacitors have played an important role in society for two centuries. The rise of the consumer electronics industry has been underpinned by innovations in battery technology, and batteries are subject to more public interest now than ever before due to the uptake of electric vehicles and the demand for grid-scale batteries to stabilise an intermittent renewable energy supply. Most electrochemical energy systems are complex. A multitude of processes over different length- and time-scales drive their realised performance. One source of complexity arises from the enigmatic mechanism by which inter-particle interactions at the atomistic scale give rise to critical macroscopic behaviours. Examples of this can be found in liquid electrolytes, for which electrostatic screening lengths and ion conductivity are both observed to deviate significantly from predictions of classical theories at high concentrations. A second source of complexity is that each system is operated in a different way in practice, and the way a system is used strongly influences both short- and long- term performance, with each system following a unique degradation trajectory. This thesis will demonstrate that both of these challenges can be tackled through the combination of data, physical intuition and machine learning. Machine learning models can learn from orders of magnitude more data than humans can, and we will see that such models can be trained to make more accurate predictions about how electrochemical systems will perform under different operating conditions. In addition, machine learning can act as a “computational microscope”, offering new ways of understanding the molecular origin of macroscopic properties. I begin by addressing the enigmatic under-screening effect observed in concentrated electrolytes, and explore whether discrepancies between classical theory and experiment can be explained using the concept of ion pairing, by identifying the number of statistically distinct environments inhabited by ions. The results bring into question the validity of the ion pair hypothesis for concentrated systems, but more importantly they suggest that static properties of electrolytes, such as screening length, can be explained by studying statistical differences between local ionic environments rather than just the mean local environment as captured by the radial distribution function. Extending this, I then posit that global dynamic properties such as ion conductivity can also be decomposed into atomistic contributions that are functions of local static structure. The idea is to learn the mapping from local structural motif to a local contribution to conductivity, which effectively generalises ideas first put forward in the ion-pair hypothesis or “cluster Nernst-Einstein” theory. By studying the distributions of local conductivities across electrolytic systems we can decipher what structural motifs are correlated with enhanced or degraded conductivity. I then turn to address the system level challenge of forecasting how electrochemical systems will respond to different operating conditions. In the growing lithium-ion battery industry, this is a major challenge, since batteries of the same chemistry will respond differently to the same use conditions due to differences in internal state caused by manufacturing heterogeneity and different extents of degradation. We develop a general framework that combines electrochemical impedance spectroscopy with machine learning to predict how a battery will respond to a given use condition, which has relevance for the design of improved battery management systems. The findings of this thesis help to further our understanding of the fundamental mechanism by which inter-ion interactions give rise to screening and conduction within liquid electrolytes. More broadly, these findings can help to guide the design of novel electrolytes for next generation systems, to develop optimal fast-charging protocols that do not sacrifice battery life, and to triage batteries towards second-life applications.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2023

Author dc:creator

• Jones, Penelope

Advisor dc:contributor.advisor

• Lee, Alpha

• battery
• conductivity
• electrochemistry
• electrolyte
• energy storage
• machine learning

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/c3f7ae00-df12-471c-93a3-79043fcbb562/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.96629

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/349809