Using machine learning to build a better battery
One of the biggest tech stories of 2016 wasn’t about awesome, new technology. It was actually about when technology goes wrong: In many ways, 2016 was the year of the exploding batteries.
A little over a year ago, hoverboards topped many holiday wish lists. By December 2015, they were being recalled by the thousands. According to Popular Science, “…cheaply made hoverboards have exploded and caught fire, forcing Amazon to stop selling specific models and Overstock to discontinue all sales.”
Rolling into 2016, major computer companies recalled batteries for fire hazards, baby monitors were pulled from shelves and major airlines diverted flights for emergency landings when tablets caught fire onboard. In September, certain smartphones were banned by airlines and added to the growing list of recalled electronics. Most recently, new reports of exploding phones have surfaced in China.
All these issues are related to problems with the lithium ion batteries, batteries that have become ubiquitous in our mobile lifestyles.
Why lithium ion batteries explode
In order to understand why lithium ion batteries are prone to exploding, it’s important to first understand how they work. These batteries are comprised of a positive electrode (anode), a negative electrode (cathode), and the electrolyte, the chemical solution between the electrodes that conducts the ions. Both the anode and the cathode are in the electrolyte, but they are separated by a physical barrier, or membrane.
The quest for smaller batteries has, in part, resulted in ever-thinner membranes separating the anode and cathode. The desire for faster charging times means that these batteries may be exposed to higher temperatures. Neither of these factors is great for battery health.
The electrolyte is typically an organic liquid, and therefore flammable. Liquid electrolytes are relatively inexpensive and conduct ions well. This has made them a popular option for consumer electronics. According to The Mirror, “… when the (extremely volatile) electrolyte is inside a sealed battery case (like in a smartphone) pressure builds up and, on rare occasions, will actually puncture the casing.”
Finding a safer electrolyte to create a safer battery
“The number of known lithium-containing solid compounds is in the tens of thousands, the vast majority of which are untested,” stated Austin Sendek, a doctoral candidate in applied physics at Stanford University. “Some of them may be excellent conductors, but there’s no way we can thoroughly measure them all in a reasonable amount of time.”
A move to a solid electrolyte could provide a safer battery. According to Sendek, “Solids are far less likely to blow up or vaporize than organic solvents. They’re also much more rigid and would make the battery structurally stronger.”
Identifying new options for electrolytes has been a research-intensive endeavor: Promising compounds are typically identified through a trial-and-error approach. Scientists have thus far invested decades in the search for a safer alternative.
Sendek worked with assistant professor Evan Reed and others in the Materials Science & Engineering Department at Stanford using artificial intelligence (AI) and machine learning to develop large-scale computational screening approach to identify promising electrolyte candidates based on existing experimental data. Their approach enables them to find potential candidates at an unprecedented speed. Their results were published in the paper Holistic computational structure screening of more than 12000 candidates for solid lithium-ion conductor materials published in the journal Energy & Environmental Science.
Out of 12,000, there were 21
The majority of the time invested in this project was spent curating known scientific data about solid compounds containing lithium. “Austin collected all of humanity’s wisdom about these materials, and many of the measurements and experimental data going back decades,” said Professor Reed, senior author on the paper. “He used that knowledge to create a model that can predict whether a material will be a good electrolyte”
The screening considered multiple additional criteria including cost and availability. The goal is to identify lithium-containing crystalline solids that demonstrate high ionic conductivity, low electronic conductivity, and both structural and chemical stability.
The training set for the machine learning algorithm included 40 crystal structures and reported experimentally measured ionic conductivity values available in the literature. It included both examples of good conductors and poor conductors.
“We used MATLAB for all the data analysis in this work: for reading in atomic structure files, extracting the features, learning the best model from the features, and then applying the optimal model to the structures in the Materials Project database,” stated Sendek, the lead author on the paper. “We used a lot of the different statistical fitting functions while we were exploring possible models for this work – least squares regression, robust regression, locally weighted least squares, SVMs, logistic regression, multiclass classification and more.”
The screening reduced the list of candidate materials from 12,831 down to 21. The computational model is approximately 5 to 6 orders of magnitude faster than experimental approaches. The speed enabled screening candidate Li-containing solids in a matter of minutes. Only a few of these have been examined experimentally, so the next step is to test the viability of actual samples using them in real world conditions. One of these 21 substances could prove to be the ideal electrolyte for lithium ion batteries!
Machine learning and material science
This research is a great example of how data science and AI are enabling smarter science. As the amount of data available increases and compute power continues to improve, this systematic approach to “scoring” materials based on previously generated data of various characteristics can be applied to a myriad of material science’s biggest challenges.
“This inefficient trial-and-error paradigm also plagues lots of different areas within materials science,” states Sendek. “Our work seeks to use machine learning to beat randomly guessing which materials will perform well.”
The group at Stanford plans to expand their machine learning efforts and in the coming years. They hope to build out a suite of robust predictive models for many different materials applications.
Check out this video to learn more about their computational screening approach to identify candidate electrolytes:
Comments
To leave a comment, please click here to sign in to your MathWorks Account or create a new one.