Global Data Science Forum

Creating a better battery with machine learning

By Tim Stone posted Thu November 12, 2020 07:20 AM


Improving batteries has been continuously hampered by slow experimentation and exploration processes. Machine learning is advancing it up by orders of magnitude.

One of the many difficulties in discovering how long any battery will be helpful is that it’s challenging to make that assessment early in its use period. That situation may vary due to a project carried out by a collaborative Stanford/MIT team operating under the auspices of Center for Data-Driven Design of Batteries. Academic-industrial cooperation involves the Toyota Research Institute and aims to integrate theory, practices, and data science.

They used AI/machine-learning elegance and algorithms to produce models that correctly predict long-term battery life by using data obtained from charge-discharge periods that’s been included only in the early steps of a battery’s life. “The usual way to test new battery designs is to charge and discharge the cells until they break. As batteries have a long life, this method can take many months and even years,” said co-lead author Peter Attia, Stanford doctoral candidate in elements science and engineering. “It’s an unreasonable bottleneck in battery investigation.”

Creating the best molecular structure blocks for battery parts is like creating a recipe for a new sort of cake, when you have billions of possible ingredients. The difficulty involves deciding which ingredients work best collectively — or, more simply, create an good (or, in the situation of batteries, a safe) output. But even with state-of-the-art supercomputers, investigators cannot accurately model every molecule's chemical properties that could prove to be the base of next-generation battery supply. 

Alternatively, researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory have applied to the power of machine learning and artificial intelligence to expedite the process of battery discovery dramatically.

As reported in two new papers, Argonne researchers first produced a highly accurate database of about 133,000 small organic molecules that could form the base of battery electrolytes. To do so, they did a computationally concentrated model called G4MP2. This set of molecules designed only a small subset of 166 billion bigger molecules that scientists needed to probe for electrolyte applicants.

Because using G4MP2 to solve each of the 166 billion molecules would have needed an impossible number of computing time and energy, the investigation team used a machine-learning algorithm to compare the precisely known constructions from the smaller data set to much more coarsely printed structures from the more massive data set.

“When it comes to discovering how these molecules operate, there are big tradeoffs among efficiency and the time it takes to estimate a result,” said Ian Foster, Argonne Data Science and Learning division manager and author of one of the articles. ​“We think that machine learning describes a way to get a molecular picture that is approximately as precise at a part of the computational cost.”

Foster and his co-workers used a less computationally taxing modeling structure based on density functional approach, a quantum mechanical modeling framework used to determine electronic structure in large ways to provide a basis for the machine learning pattern. Solidity functional theory presents a good approximation of molecular properties but is less reliable than G4MP2.

Improving the algorithm to ascertain better knowledge about the broader class of organic units involved comparing the molecules' atomic positions measured with the highly accurate G4MP2 versus those investigated using only density practical theory. By using G4MP2 as a gold example, the researchers could train the density practical theory model to include a correction factor, improving its accuracy while keeping computational prices down.

“The machine learning algorithm provides us a way to look at the connection between the atoms in a large molecule and their friends, to see how they bond and communicate, and look for connections between those molecules and others we know very well,” said Argonne computational scientist Logan Ward, a writer of one of the studies. ​“This will assist us to make forecasts about the energies of these larger particles or the differences among the low- and high-accuracy predictions.”

“This whole scheme is designed to give us the most important picture possible of battery electrolyte competitors,” added Argonne chemist Rajeev Assary, an inventor of both studies. ​“If we are continuing to use a molecule for energy storage purposes, we need to know features like its stability, and we can use this machine learning to divine properties of bigger molecules more carefully.”

Extra collaborators on the studies involved Argonne’s Ben Blaiszik, Larry Curtiss, and Paul Redfern and Badri Narayanan of Argonne and the University of Louisville.

Funding for the project was provided by the Joint Center for Energy Storage Research, a Department of Energy Innovation Hub. The researchers used sources of the BEBOP supercomputing cluster at Argonne’s Laboratory Computing Resource Center.

Unleashing AI on battery construction is good news for a warming world. Battery accommodation is a key factor in improving the amount of renewable power on the grid, and when it comes to decarbonizing our power equipment, time is of the essence. After decades of plodding process, AI-driven battery research undertakes to pick up the pace finally. “This is all tied back to decarbonization,” says Chueh. “We want to get there immediately because we don’t have much chance left.”

Smart ideas for cost-efficient battery making

A broad range of elements must be taken into account, starting with the selection of locations. For instance, what are the technical support and logistical requirements at a particular site? What about the availability and price of raw elements? Is the area engaging for well-trained and qualified staff — and are there enough people to pick from in the local workforce? How about the tax status? What do everything look like when it occurs to political resistance and regulation? The more solidly investigated the answers to such problems, the better the choice of location will be.

The most crucial thing is that the China custom ev lithium ion battery factory idea is developed with an knowledge of technology and methods — something that needs extensive expertise and solid know-how. Finally, the purpose is to create a full package, from place, and the building's planning to the factory's technical support and development.