AI prescribes new electrolyte additive combinations for enhanced battery performance

Batteries, like humans, require medicine to function at their best. In battery technology, this medicine comes in the form of electrolyte additives, which enhance performance by forming stable interfaces, lowering resistance and boosting energy capacity, resulting in improved efficiency and longevity.

Finding the right electrolyte additive for a battery is much like prescribing the right medicine. With hundreds of possibilities to consider, identifying the best additive for each battery is a challenge due to the vast number of possibilities and the time-consuming nature of traditional experimental methods.

Researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory are using machine learning models to analyze known electrolyte additives and predict combinations that could improve battery performance. They trained models to forecast key battery metrics, like resistance and energy capacity, and applied these models to suggest new additive combinations for testing.

By combining machine learning with experimental testing, researchers quickly identified effective electrolyte additives, accelerating the discovery process compared with traditional methods, which are costly as well as time-consuming.

This research, now published in Nature Communications, successfully found new additive combinations that outperformed existing ones, showing the power of data-driven techniques in advancing battery technology and paving the way for high-performance, efficient batteries.

Prescription for peak performance

LiNi_0.5Mn_1.5O₄ batteries—composed of lithium, nickel, manganese and oxygen, known as LNMO—operate at a high voltage and offer significant advantages to traditional batteries. They have a higher energy capacity and eliminate the need for cobalt, a critical material associated with supply chain concerns.

While the higher voltage of LNMO batteries offers benefits, it also presents significant challenges. Cellphone batteries and individual electric vehicle cells typically operate at low voltage, around 4 volts. But an LNMO battery operating at 5 volts far exceeds the stability limit of any known electrolyte.

“High voltage usually indicates high energy density,” explained Chen Liao, an Argonne chemist and senior scientist at the University of Chicago. “But it also presents numerous challenges because the electrolyte and cathode are in a highly energized state that can lead to decomposition. Operating at such a high voltage can be both a blessing and a curse—the battery materials must be exceptionally stable.”

Introducing an electrolyte additive to the LNMO battery could help limit decomposition and improve battery performance. The researchers found that the ideal additive decomposes during the first few battery cycles, forming a stable interface on both electrode interfaces.

This layer helps lower resistance, which means less energy is wasted and less degradation occurs, boosting the battery’s energy output. Using an additive is also an economic approach. Battery manufacturing processes are mature and unlikely to change but simply adding an additive to the electrolyte formulation is a straightforward change to adopt.

“Think of an additive like medicine,” Liao said. “It makes the battery better.”

Making connections with machine learning

To efficiently and affordably explore the extensive realm of chemical possibilities, scientists are using machine learning techniques for discovering and optimizing materials. These techniques allow for predicting material properties, designing material structures with desired functionalities and identifying material candidates through dataset analysis.

Liao, an experimentalist, teamed up with Hieu Doan, a computational scientist at Argonne, to develop a machine-learning model to explore possible electrolyte additives and determine their effect on LMNO battery performance.

“The ultimate goal of this work was to quickly screen for the best additive for the system,” Doan said. “These additives are organic molecules with different chemical structures, so they come in different shapes and size. The challenge was how to look at their chemical structure and predict their performance.”

To develop this model, they needed to collect initial data but were limited by the number of experiments that could reasonably be performed. Instead, they focused on creating a diverse initial dataset of 28 additives that incorporated various functionalities to train the model effectively.

This approach ensured that the model could recognize various functionalities during training, enabling it to make accurate predictions in the future. To develop a machine-learning model capable of predicting the performance of battery additives, the researchers needed to “map” the chemical structure of each additive to its performance within the battery system. They achieved this mapping by examining the features of the additive molecules, known as descriptors.

Doan explained, “How can we describe these molecules so that we can use the descriptor to make a prediction on performance?” He likened this process to inferring someone’s profession based on their appearance; for instance, someone wearing a suit and carrying a briefcase might be assumed to be a lawyer.

“Based on that feature, you make that connection. You’ve seen that before from experience and you correlated those two things together,” Doan said.

The machine learning model is designed to follow a similar logic, establishing a connection between the chemical structure of additives and their impact on battery performance, much like how humans make connections based on experience.

Predicting success

After training the model using the initial 28 additive dataset, Liao and Doan were able to predict the performance of 125 new combinations of additives. The model successfully identified several promising additives that improved battery performance, outperforming additives from the initial data.

This method not only saved time and resources but also demonstrated how machine learning can accelerate the discovery of new materials with desired properties for better batteries. By avoiding 125 traditional experiments, which would have taken approximately four to six months and required significant equipment costs, the researchers showed how machine learning can streamline discovery using a small experimental dataset.

“The traditional idea is that you need a lot of data to train a machine learning model,” Doan said. “But our work shows that you don’t need a lot of data to train an accurate prediction model. You just need a good set of data to do it properly.”

By finding the right “prescription” through machine learning, scientists can ensure batteries operate at their best, paving the way for more efficient and longer-lasting energy solutions.