Using AI to identify battery electrolytes

Dr. Neil Canter, Contributing Editor | TLT Tech Beat March 2020

Enthalpy of formation was determined to assess molecular stability.

---

Figure 3. In an effort to facilitate the identification of more effective and stable battery electrolytes, research has been conducted to identify candidates through the use of machine learning and artificial intelligence. Figure courtesy of Argonne National Laboratory.

---

KEY CONCEPTS
• Instability is the current types of batteries under development has been traced to the organic electrolytes used. 
• Identification of more stable electrolyte candidates was accomplished by evaluating the enthalpy of formation for 133,000 organic molecules in a specific database.
• High-performance computing and AI were used to identify specific candidates. 

Extensive research is ongoing to develop a commercially viable battery that is a cost- effective alternative in automobiles to the internal combustion engine. The leading candidate continues to be lithium-ion batteries.

But lithium-ion batteries face safety issues that are caused by instability traced to the organic electrolytes used which can facilitate the formation of dendrites. If left unchecked, dendrites can grow between the anode and cathode, leading to short circuiting of the battery which can cause overheating and eventually a fire. In a previous TLT article¹, researchers discussed the development of a solid electrolyte based on aramid nanofibers and poly (ethylene oxide) that appears to stop the growth of dendrites.

As an approach to better understand how lithium-ion batteries function and to expedite their development researchers have turned to artificial intelligence. A previous TLT article² reported on a study to predict the life of lithium-ion batteries using AI. Researchers initially established a data set based on results from 124 commercial fast-charging batteries that underwent varying charge/discharge cycles. Ultimately, the researchers compiled a data set with approximately 96,7000 cycles that predicts the operating life of specific lithium-ion batteries. Of the models used, the most accurate one to predict battery life focused on changes in voltage as a function of the number of cycles.

Dr. Rajeev Assary, chemist at the U.S. Department of Energy’s Argonne National Laboratory in Argonne, Ill., says, “Organic molecules are contained within the electrolyte of a lithium-ion battery or the next-generation redox flow batteries. Three parameters that must be used to evaluate their performance include oxidation-reduction (redox) window, solubility and stability. The latter is very important as it will directly determine the operating lifetime of the battery.”

In evaluating electrolytes, Assary indicated that predicting molecular stability is very important. To accomplish this task, thermodynamics is a useful tool. Assary says, “Evaluating accurate enthalpy of formation values is essential to screen large chemical space including the measure of molecular stability. This parameter is a measure of the enthalpy change that occurs when a molecule decomposes to its pure elements.”

Assary and his colleagues have now determined the enthalpy of formation for 133,000 small organic molecules that contains the molecular fingerprints of the battery electrolyte candidates using High Performance Computing and AI (see Figure 3).

GDB-9 database
The researchers initially examined a database known as GDB-17. Assary says, “We wanted to work with the GDB-17 database because it contained a total of 166 billion organic molecules and would have given us insight into a large number of options. GDB-17 contains all molecules of up to 17 heavy atoms of the first row of the periodic table (including carbon, nitrogen, oxygen and fluorine) and hydrogens. Unfortunately, there is insufficient computing power to utilize GDB-17.”

Instead, the researchers evaluated a subset of 133,000 organic molecules that comprise GDB-9 where the number of first row atoms is reduced to nine. Assary says, “In our work, we used molecules that contained carbon, nitrogen, oxygen, sulfur and some but not too many halogens.”

The researchers used a computationally intensive and accurate model known as G4MP2 to assist with calculating the enthalpy of formation. Assary says, “G4MP2 is based on a classic theory that evaluates all electronic interactions. It provides an accurate analysis of the energies of the organic molecules in GDB-9.”

At the same time, the researchers used a less accurate computationally based modeling framework that is derived from density functional theory, which is a quantum mechanical modeling framework that calculates the electronic structure of in large systems. The objective was to provide a basis for the machine learning model.

To facilitate the process, the researchers selected 459 organic molecules that have known enthalpies of formation to use in the machine-learning process. Assary says, “We obtained good correlation (less than 1 kcal per mole) between the experimentally determined and G4MP2 determined enthalpies of formation for these molecules.”

The researchers also used G4MP2 and a less accurate density functional theory method known as B3LYP to evaluate the enthalpies of formation for a small set of 66 organic molecules containing between 10 and 14 heavy atoms. Assary says, “We compared experimental values with computational values and also found good correlation that was within 1-2 kcal per mole.”

The next objective for the researchers is to use G4MP2 and density functional theory to evaluate molecules with charges. Assary says, “We have initially worked with neutral molecules, but charged molecules and fragments are present in battery electrolytes, and this must be considered in selecting the proper one. Empirical electrochemical techniques such as cyclic voltammetry will be used to compare values obtained through machine learning.”

The researchers hope to predict at what voltage a specific molecule will oxidize or reduce. Assary says, “We also intent to use solubility as a criterion to identify electrolyte candidates.”

Additional information can be found in two recent articles^3,4 or by contacting Assary at assary@anl.gov. 

REFERENCES
1. Canter, N. (2015), “Dendrite-suppressing battery technology,” Tribology & Lubrication Technology 71 (4), pp 14 – 15
2. Canter, N. (2019), “Predicting life of lithium-ion batteries,” Tribology & Lubrication Technology 75 (7), pp 12 – 13

Neil Canter heads his own consulting company, Chemical Solutions, in Willow Grove, Pa. Ideas for Tech Beat can be submitted to him at neilcanter@comcast.net.