Is Machine Learning ready to tackle the “AND problem” of Batteries?

Venkat Viswanathan
5 min readOct 18, 2021

October 15th marked my one-year anniversary as Chief Scientist at Aionics Inc. This gives me a chance to reflect on my assumptions from a year ago on whether machine learning is ready to tackle challenges in practical battery materials design and optimization. Those that know me know that I take a long time to be convinced to act on something, and when I do, I try to go full steam. I’d like to reflect on the journey that led me to join the Aionics ride and the questions around machine-learning guiding battery design.

As a graduate student, I had worked on lithium-air batteries and I summarized the state of the field as “the good, the bad and the ugly” in 2013. The bad was the challenge with the liquid electrolyte which would react in the harsh environment of the lithium-air battery. Liquid electrolytes consist of a lithium salt dissolved in an organic solvent, and given the vastness of organic chemistry, I thought there was hope in finding a liquid electrolyte that would work for lithium-air batteries.

When I started as an Assistant Professor in 2014, I decided that my angle to this had to be data-driven. Organic chemistry is nuanced and there are as many exceptions as there are rules. With this vision, I launched SEED, System for Electrolyte Exploration and Discovery, an internal tool that we put to use to design liquid electrolytes. Although launched with aim of finding electrolytes for lithium-air batteries, community interest had dwindled and I transitioned to working on lithium-metal batteries.

Lithium metal anodes have been the holy grail for batteries, as I have written earlier. To enable lithium metal anodes, I took two shots on goal, (i) a solid electrolyte approach, that I have talked about earlier, and (ii) a liquid electrolyte approach, that would self-form a “solid electrolyte”. We were well on our way to using data-driven approaches for the second approach. The solid electrolyte approach, on the other hand, was data-starved.

Here comes the first major event in the story. In late 2016, Austin Sendek, then a PhD student at Stanford, was bold enough to train a data-driven model with only 40 data points to screen over the known inorganic lithium-containing chemical space. The conventional wisdom and probably most PhD advisors (including me) would have told him not to do this. I was intrigued by the paper and asked Austin to present a seminar over videoconference to my group. This inspired us to jump in and begin using machine learning to screen for solid electrolytes.

Despite these methods’ great success in academia, I did not think that they could find utility in industry. The second important twist here was my conversations with Tim Holme, CTO of QuantumScape. We both independently came to the conclusion that good statistical inference paired with simple machine learning has the potential to rapidly accelerate the rate of battery materials R&D. I decided to teach a new course at Carnegie Mellon University, called Bayesian Machine Learning and Tim agreed to co-develop the class. The course was offered in 2019, and it required a steep learning curve for engineering students. Tim delivered guest lectures and we had built incredible content, which if put to practice in industry held promise.

To appreciate the challenge, it is useful to understand the time-to-market for materials innovation. It takes on average about 18 years to transition a lab innovation to a commercial product. A significant portion of that time is spent in optimizing and scaling manufacturing, and if that could be shortened, this would represent a step change. In 2019, my then PhD student (now at Tesla), Vikram Pande, who had helped develop SEED, used low-data physics constrained ML models for designing liquid electrolytes. By then, SEED tool had several success stories in reducing time to discovery in an academic setting. He told me that Austin had spun out Aionics Inc, aiming to take a shot at this challenge in industry. I thought it was not ready for two reasons — the methods were not quite there for noisy field data and the mindset in companies needed a change to adopt this.

Later that year, I was awarded an ARPA-E DIFFERENTIATE award to scale up machine learning methods for electrochemical systems, together with incredible partners, Julia Computing, Citrine, and MIT. ARPA-E programs enable (and force) the PIs to scale the technical efforts as well as gain a pulse of the market needs. I was slowly leaning towards the possibility that maybe these methods were ready after all.

The grand challenge in data-driven material design for batteries is to be able to identify a component material (e.g. cathode, anode, electrolyte), that can address the “AND problem” of batteries. Batteries have to satisfy many requirements simultaneously, and mapping the device requirements to the battery component, e.g. liquid or solid electrolyte, is a challenge. The main goal in the phase of invention and optimization is to directly map material identity (e.g. identity of organic solvent, salt concentration, etc) to device performance (e.g. cycle life, fast-charging capability, etc).

On July 29, 2020, I tuned in to an evening webinar hosted by Young Professionals in Energy: Boston, (I still don’t know how I came across this!), where Austin Sendek, CEO of Aionics Inc., presented a plot that I remember vividly. He showed a direct mapping from liquid electrolyte identity to battery performance. The most intriguing part about this was that it was not on clean data from an academic setting but on the field with (noisy?) customer data. I was immediately convinced and I wrote an email to Austin that evening to discuss a joint path forward and on October 15th, I joined as Chief Scientist.

The first client project I worked on was with Sepion Technologies. They were starting a new project, led by Jessica Golden, and we were able to jointly define the metrics, project testing plan, etc, from the beginning. The major challenge in practice is to modify the current practices to be ready for data-driven design. A fresh start meant that we could do things right, and in a short six-month period, we had built a joint pipeline that was eliminating unwanted experiments efficiently and innovating at a rapid pace. Along the way, at Aionics, we developed a DFT product offering that can more confidently decide on candidates that have never been tested before.

A year since joining, it is abundantly clear to me that this is the path forward. Every battery materials company needs a co-innovation partner; it is not a nice-to-have, but a need-to-have to compete with the rate of progress necessary to meet the aggressive electrification goals and more importantly, address the climate change challenges ahead. Two prominent examples of this co-innovation model are (i) QuantumScape working with Landing AI to bring deep-learning based visual inspection to improve materials quality, and (ii) Chement Inc., a company I co-founded with Breakthrough Energy Fellow, Greg Houchins, is partnering with Aionics Inc., before even performing their first experiment. We are just at the beginning and there is a long way to go to reach the success levels of drug discovery, but it shows that using data-driven methods to truly optimize a new battery material/chemistry is simply a matter of when, not if!

--

--

Venkat Viswanathan

Associate Professor @CarnegieMellon University, Advanced Batteries, Electrochemical Devices