Aionics has provided customers small quantities of AI-tuned battery electrolyte for testing.
There are many requirements for electric air taxi batteries: high power for vertical takeoff, high energy for cruise, fast charging for quick turnaround, long cycle life for operating economics and safety for certification. Artificial intelligence is being applied to optimize these often conflicting needs.
In recent years, lithium-ion battery development has been focused on electrode materials that increase energy density for greater electric vehicle range. Now that the range challenge is largely conquered, attention has shifted to other battery attributes, such as safety, cost, power, charging and cycle life.
- Electrolyte determines performance attributes
- Drop-in component can tune batteries for different markets
At the same time, batteries are finding new applications, from grid storage to electric aircraft, that require different mixes of these attributes—mainly determined by the electrolyte, a chemical solution that enables lithium ions to shuttle between the cathode and anode.
Researchers are taking a leaf from the drug discovery playbook and using artificial intelligence (AI) to analyze combinations of 10 billion available molecules in search of electrolyte formulations tuned to different customers’ specific requirements, whether for high power, long life or both.
Californian startup Aionics was formed in 2020 by three postgraduates from Stanford University. “We were working at the intersection of materials science and AI,” co-founder and CEO Austin Sendek says. “I was interested in the potential for accelerating the design of materials and especially batteries.”
The seed-stage company based in Palo Alto, California, has secured investment from mobility-focused Trousdale Ventures and UP.Partners along with the University of Michigan’s startup fund and Avila VC. Aionics has signed joint development partnerships with Porsche’s battery subsidiary Cellforce Group and more recently with an undisclosed aviation company to “tweak” an off-the-shelf cell into an electric vertical-takeoff-and-landing (eVTOL) battery.
“The idea has been to discover and commercialize new battery materials,” Sendek says. “We’re focusing on optimizing electrolyte formulations around the various performance profiles that applications need.” Aionics is mainly working with liquid electrolyte systems, the type most widely used in the battery industry.
“Every application area, whether it’s eVTOL, automotive or grid-scale storage, needs something slightly different,” he adds. “And batteries today are kind of one-size-fits-all.
“Reengineering the electrolyte can unlock a lot of changes in performance along with different unique properties that are determined by the electrolyte,” Sendek explains. “And the marginal cost to change the electrolyte is close to zero. It’s a drop-in component. You don’t have to retool your factory. You can explore tons of combinations relatively easily, and the design space is huge.”
The majority of electrolytes used in commercial batteries are different ratios of the same 11 molecules. “If you want more power, you add a little bit more of this, and if you want a little bit longer lifetime, then you can get a little bit more of that,” Sendek says.
“But that’s a tiny fraction of chemical space that’s available,” he notes. “There are about 10 billion molecules you can order and have show up in a couple weeks. We’re searching through the same chemical design space that drug companies are looking through for therapeutics. And essentially any molecule you can buy that is stable—including some that are produced by generative AI and are feasible but have never been made—is a candidate for this process.”
Aionics uses four or five layers of AI. “If I want to optimize the cell to give me high power for 5 min., I need to build a machine-learning model that can predict, based on formulation of the electrolyte, what the peak power of the cell is going to be,” Sendek says. “And I need to apply that really specific model across this chemical space of 10 billion possible molecules. So it needs to be computationally fast, but it also needs to be accurate.”
Such models can use a deep-learning potential neural network function that simulates interaction between atoms. This type of model predicts the potential energy of a system and is used in molecular simulations. Aionics also builds small data models that learn the response of the cell to changes in the electrolyte. “That’s like a regressor-type model,” he says.
Aionics uses a multimodal approach to data collection that combines large language models trained on textbooks, publicly available databases, simulations and experimental data from partner laboratories and customers. “I call this a semi-empirical approach,” he says. “There’s bottom-up science, and there’s top-down data, and you want to find some way to bridge that.”
Aionics has produced small amounts of electrolyte formulations that are being tested by customers. “Our commercialization strategy is not to become a manufacturer but to find companies with the right manufacturing capacity for the formulations,” he says. “We’re in the Pfizer-BioNTech model.”
Related Content
