It would be great if the battery companies could say which of their batteries will last for at least two years and sell them to mobile phone producers, and which will last for ten years or more, and sell them to electric vehicle suppliers. The new cooperative study released in Nature Energy demonstrates how they could achieve this goal.
By combining extensive experimental information and artificial intelligence techniques, Scientists at the Massachusetts Institute of Technology (MIT), Stanford University and the Toyota Research Institute (TRI) found the key to correctly anticipating the useful life of lithium-ion batteries before their capabilities began to decline. After the researchers trained their machine learning model with a few hundred million data points, based on the decline in voltage and a few other factors in the early discharge cycles, the algorithm predicted how many more cycles each battery would last. The projections were within 9% of the real life cycle. Separately, the algorithm categorized batteries as either long or short life expectancy based on only the first five charging/discharge cycles. Here, 95 percent of the time, the predictions were correct.
This machine learning technique could accelerate the research and development of fresh battery models and decrease time and manufacturing costs, among other apps. Researchers have made the data accessible to the public — the biggest of its kind.
The normal way to test a fresh battery design is to charge and discharge the cells until they die. Since the batteries have a lengthy lifespan, this method can take months and even years. It’s a costly bottleneck in the field of battery studies.
Work has been carried out at the Center for Data-Driven Design of Batteries, an academic-industrial partnership that integrates theory, experiments and data science. The Stanford scientists, led by William Chueh, Assistant Professor of Materials Science & Engineering, performed battery experiments. The team of MIT, led by Richard Braatz, Professor of Chemical Engineering, carried out machine learning research. Kristen Severson was also the co-leader of the studies. She finished her Ph.D. in Chemical Engineering at MIT.
One of the key functions in data-driven, multi-institute research projects is to ensure that big data streams generated at experimental installations are managed and effectively transmitted between distinct study organizations. Study co-authors Muratahan Aykol and Patrick Herring have brought the experience of TRI with big data to the project and their own expertise in battery development to enable the efficient management and seamless flow of battery data, which was essential for this collaboration in the creation of accurate machine-learning models for the early prediction of battery, fail.
Optimizing the fast charging process
One focus of the project was to discover a better way to charge batteries in ten minutes, a function that could accelerate the mass acceptance of electric vehicles. To produce the experimental data set, the team loaded and discharged the batteries until each of them had reached the end of their helpful lives, which they described as a 20 percent loss of their maximum ability. In the process of optimizing rapid charging, the scientists wanted to figure out whether they needed to operate their batteries on the ground. Advances in computational power and information generation have lately allowed machine learning to accelerate the computational progress on a variety of tasks. These include the forecast of the characteristics of materials. Their findings demonstrated how far into the future the conduct of complex systems can be predicted.
The capability of the lithium-ion battery is generally stable for a while. Then it requires a sharp downward turn. The plummeting point differs extensively, as most customers of the 21st century understand. In this project, the batteries lasted from 150 to 2300 cycles. This variance was not only partially due to the testing of distinct quickly charging techniques, but also due to the ordinary variations that arise in commercially manufactured instruments that rely on molecular interfaces.
According to Attia, the new technique has many prospective applications. For instance, the time needed to validate batteries with fresh chemistries can be shortened. This is particularly crucial considering the fast developments in plastics. Manufacturers can also use the classification method to grade batteries with longer lifespans to be sold at greater rates for more challenging uses, such as electric vehicles. Recyclers can use the technique to discover cells in used EV battery packs that have enough life in them for secondary use. The last step in the production of batteries is called “formation” which can take days to weeks. Using this approach could significantly reduce and lower the cost of production.
Researchers are using this early prediction model to optimize charging procedures that could allow batteries to be charged in ten minutes. Using this model, the optimization time can be reduced by more than a factor of 10, significantly accelerating research and development. This research is part of the Accelerated Materials Design and Discovery (AMDD) program of TRI. Led by program director Brian Storey, the $35 million initiative works with research institutions, universities and companies to use artificial intelligence to accelerate the design and discovery of advanced materials.
The combination of ample experimental data and AI revealed the key to accurately predicting the useful lifetime of LIPO batteries before their capacities began to decline. Once the researchers managed to train their machine learning model with charging and discharging data points of a few hundred million batteries, based on the voltage decline and some other factors in the early cycles,
the algorithm was able to accurately predict how many more cycles each battery would last. The new method has many potential applications, such as shortening the time needed to validate new types of batteries, which is particularly important given the rapid advances in materials. Using the classification technique, electric vehicle batteries designed to have short lifespans–too short for cars–could instead be used to power street lights or back up data centers. Recyclers were able to find cells from used EV battery packs with enough capacity left for a second life.
This groundbreaking AI implementation could optimize the production of batteries by reducing production time and production costs. This important AI application is also promoted and supported by AI World Society (AIWS) to develop an advanced AI technology for optimizing production and improving the quality of a better human society.