Applying Machine Learning Techniques to Lithium-ion Cell Research

Abstract

Progress in lithium-ion cells research is largely a matter of determining which aspect of the cell’s design and operation will lead to longer life, higher energy-density, and lower costs. These attributes can easily be characterized empirically but determining if a cell will last 10 or 20 years in the naïve way is much too slow (i.e. it would take between 10 to 20 years). The same attributes could in principle be determined from theory alone, but this is a very challenging problem and is currently unsolved.

It would therefore seem that the way forward is to leverage some experimental results obtained in a reasonable time to estimate the key attributes of various cell designs and make progress towards better designs. The development of models and tools using data (i.e. machine learning) is a powerful and much studied toolbox which in principle is ideally suited to the task at hand, but care must be taken in its application. If this thesis withstands the test of time, it will likely do so as the initiation of a process of cross-pollination of the field of machine learning towards lithium-ion cells research. To this end, we offer two clear applications of machine learning to the understanding of specific measurements. We apply machine learning to impedance spectroscopy and then to Fourier-transform infrared spectrometry. Finally, we offer an example of a data processing system scaled to support the long-term cycling data of a laboratory in the real-world.

As such, it is our hope that the process of cross-pollination will be helped by these concrete in-depth examples of applying the techniques of machine learning, and by a scalable system which organized tens of thousands of experiments to be used for future inquiry.