Machine Learning for Accelerated Discovery of promising Battery Materials

< back to research overview

By unleashing machine learning algorithms on the last several decades' worth of data:

  • we have predicted the performance of every known lithium-containing inorganic crystalline solid as a solid-state electrolyte in a lithium ion battery
  • we are working on identifying novel chemistries for low chemical expansion - high voltage cathodes
  • we are investigating chemical stability of lithium anodes in solid state batteries

Batteries will play an increasingly important role in society as our electronics become increasingly portable and as our transportation systems move away from fossil fuel-based energy sources. Although significant advances have been made in batteries over the last several decades, many issues remain, including concerns over cycle life and safety. There are many ways to improve the current standard of Li-ion batteries, and we utilize computational methods to identify new materials predict properties, and gain insight into electrochemical performance of novel battery chemistries.

Conductive Solid Electrolytes

By replacing the flammable liquid electrolyte of conventional batteries with a stable solid, we can drive significant improvements in battery safety, lifetime, and energy density. But solid batteries are complicated systems - and finding new electrolyte materials that can stand up to the many different demands of battery performance is no easy task. Candidate materials must exhibit high lithium conductivity, robust chemical and phase stability, a wide electrochemical stability window, low electronic conductivity, and low cost. The past several decades have seen many studies of candidate materials for solid-state electrolyte applications - but these studies typically focus only on ionic conductivity and candidates are chosen for study on a guess-and-check basis. For the first time, we have aggregated the available data from past studies and analyzed it though the lens of machine learning to drive new design insights. This allows us to strategically select for the most promising materials rather than continue in trial-and-error.


For every stable, nonmetallic lithium-containing material, we predict the likelihood of exhibiting fast lithium conduction versus the level of model extrapolation required to make the prediction. We seek materials in the upper left-hand corner of the plot.

Ionic conductivity is generally thought to be the most restrictive constraint on candidate materials, and is also the most time consuming to evaluate computationally or experimentally. We compile data on 40 materials with both good and bad measured room temperature lithium conductivity values. We then “show” these examples to a logistic regression classifier, which “learns” to predict whether that material performed well or not based on the atomistic structure. Once this training is complete, we unleash the trained model on the more than 12,000 lithium-containing solids, and find that around 1,000 of them have a >50% chance of exhibiting fast lithium conduction. To shrink the space of candidate materials further, we also predict the performance on a host of other metrics, including stability, electronic conduction, cost, and earth abundance. With all metrics considered, we are able to screen 12,000 candidates down to the 21 most promising structures in a matter of minutes. This work represents the first attempt to screen all known lithium-containing solids for electrolyte candidates with a holistic screening criteria, and because of the machine learning approaches employed we do it approximately one million times faster than evaluating each material one-by-one.

Novel Cathode Chemistries

The current standard for a positive electrode is a layered structure that allows for intercalation of lithium during cyclic charging and discharging of a battery. A transition metal with multiple oxidation states easily absorbs or gives charge to the intercalating lithium, while an oxide or phosphate lattice keeps the structure of the cathode. A successful cathode material has a high voltage against Lithium, a high capacity, high energy density, all while keeping as consistent a volume as possible during intercalation. Many research efforts have focused on optimizing one of these properties, but it is difficult to balance for example, because a high capacity means morecharge transfer which means materials must expand to accommodate more atoms. Finding cathode materials with lower chemical expansion would allow more mechanically stable batteries, less likely to crack due to huge volume changes. Even across a set of 1300 materials for which we have electrochemical data, a limited number perform well in all of the categories ofelectrochemical performance, because our standard is to find materials which outperform currentindustry standards. We hope to utilize our computational methods to investigate large scale screening into new materials spaces, giving insight to unexplored cathode chemistries.

Interfaces

Our newest endeavor involves putting these battery pieces together. While we identify new chemistries and structures of electrolytes and cathodes, we also check to make sure that they are compatible. In an ideal solid-state battery, we could use the highest energy density pure lithium anode. In reality, lithium is extremely reactive, and we are currently investigating realistic electrolyte candidates that fit all of the necessary requirements to be a great electrolyte, with the next step of ensuring that the electrolyte would work at an interface with a lithium anode. This is a new field, but with computational methods such as molecular dynamics, we are able to investigate the exact reactions at the interface and determine the time and extent of any degradation that occurs.

Contact: Brandi Ransom, bransom_at_stanford.edu

Publication:

Sendek, A. D., Antoniuk, E. R., Cubuk, E. D., Ransom, B., Francisco, B. E., Buettner-Garrett, J., Cui, Y., and Reed, E. J., Combining Superionic Conduction and Favorable Decomposition Products in the Crystalline Lithium-Boron-Sulfur System: A New Mechanism for Stabilizing Solid Li-Ion Electrolytes. ACS Appl. Mater. Interfaces(2020).

Sendek, A. D., Cheon, G., Pasta, M. and Reed, E. J., Quantifying the Search for Solid Li-Ion Electrolyte Materials by Anion: A Data-Driven Perspective. The Journal of Physical Chemistry C 124 (15), 8067-8079 (2020).

Sendek, A. D., Antoniuk, E. R., Cubuk, E. D., Francisco, B. E., Buettner-Garrett, J., Cui, Y., and Reed, E. J., A New Solid Li-ion Electrolyte from the Crystalline Lithium-Boron-Sulfur System. JOULE-D-19-0045 (2019).

Cubuk, E. D., Sendek, A. D., Reed, E. J., Screening billions of candidates for solid lithium-ion conductors: A transfer learning approach for small data. The Journal of Chemical Physics 150 (21), 214701 (2019).

Sendek, A. D., Cubuk, E. D., Antoniuk, E. R., Cheon, G., Cui, Y., Reed, E. J., Machine learning-assisted discovery of solid Li-ion conducting materials. Chemistry of Materials 31 (2), 342-352 (2019).

Sendek, A. D., Yang, Q., Cubuk, E. D., Duerloo, K.-A. N., Cui, Y., Reed, E. J., Holistic Computational Structure Screening of more than 12,000 Candidates for Solid Lithium-ion Conductor Materials. Energy & Environmental Science, doi:10.1039/C6EE02697D (2016).

News coverage:

Using machine learning to build a better battery. MathWorks Behind the Headlines Blog, January 8, 2017. Join the Facebook discussion here and here!

Stanford Researchers Work On Solving Lithium-Ion Battery Explosions. CBS SFBayArea, December 19, 2016

No more burning batteries? Scientists turn to AI to create safer lithium-ion batteries. Stanford Precourt Institute for Energy News, December 15, 2016