Artificial Intelligence and Experimental Validation Reveal the Atomic-Scale Basis for Improving the Performance of “Water-in-Salt” Batteries<\/p>\n\n\n\n
Upton, New York\u2014A team of scientists from Brookhaven National Laboratory and Stony Brook University have used artificial intelligence (AI) to gain insights into how zinc-ion batteries work and explore ways to improve their efficiency to meet future energy storage needs. Their findings, published in the journal *PRX Energy*, focus on the water-based electrolyte responsible for transporting charged zinc ions during charging and use. The AI \u200b\u200bmodel analyzed how these charged ions interact with water in solutions of zinc chloride (ZnCl\u2082, a water-soluble salt) at different concentrations.<\/p>\n\n\n\n
TheAI<\/strong> \u200b\u200bfindings were experimentally validated at the National Synchrotron Radiation Facility II (NSLS-II) at Brookhaven National Laboratory, demonstrating why high salt concentrations produce optimal battery performance.<\/p>\n\n\n\n Deryu Lu, a scientist in the Theoretical and Computational Group at the Center for Functional Nanomaterials (CFN) at Brookhaven National Laboratory who led the research, explained that zinc-ion batteries, like all batteries, convert the energy generated by chemical reactions into electrical energy.<\/p>\n\n\n\n Lu and his collaborators knew that previous research had found that the water-breaking reaction was inhibited in a special zinc chloride electrolyte. This electrolyte had a very high salt concentration and was called a \u201cwater-in-salt\u201d electrolyte, in contrast to the more common \u201csalt-in-water\u201d electrolyte. To investigate why the high-salt electrolyte was superior, they wanted to capture the atomic-scale details of how zinc and chloride ions move and interact with water at different salt concentrations, and how this interaction affects the electrolyte’s conductivity.<\/p>\n\n\n\n However, observing these atomic-scale details is extremely difficult. Therefore, the research team turned to a computer modeling method enhanced by artificial intelligence vision.<\/p>\n\n\n\n Developing AI<\/strong> Vision<\/p>\n\n\n\n Professor Lu stated, \u201cThese complex details cannot be observed using traditional computing techniques. Traditional simulation methods cannot handle the large number of atomic interactions with the required precision, thus failing to capture the timescale of evolution in such systems. Such calculations require enormous computing power and can easily take years.\u201d<\/p>\n\n\n\n Therefore, instead of performing all the complex calculations required to simulate ion-water interactions, the research team used traditional simulation methods to generate a small amount of simulated data (called a \u201ctraining set\u201d) and fed it into the AI<\/strong>I \u200b\u200bprogram. They utilized computing resources at the Theory and Computation Facility (CFN) (a user facility of the U.S. Department of Energy\u2019s Office of Science) and the Scientific Computing and Data Facility (CDS) within Brookhaven National Laboratory\u2019s Division of Computation and Data Science (CDS).<\/p>\n\n\n\n \u201cWe need a small amount of data, collecting data by computing a small number of interactions to initiate the initial model training process,\u201d said Cao Chuntian of CDS, the paper\u2019s first author. \u201cThen, we run the model to generate more data, continuously improving the model\u2019s predictive capabilities.\u201d<\/p>\n\n\n\n At each step, the scientists fed the results into a set of machine learning (ML) models to evaluate the accuracy of the predictions. Lu likened the process to calling several friends and asking them to help answer questions from the once-popular TV game show “Who Wants to Be a Millionaire?”. “If the friends\/models all agree, then your prediction is likely accurate,” he noted.<\/p>\n\n\n\n But as Cao pointed out, “When we find that some predictions in the machine learning model ensemble are significantly biased, we re-perform traditional calculations to get the correct answers. Then, we add these new corrected data points back into the training data to further improve the machine learning model.”<\/p>\n\n\n\n This iterative “active learning” process minimizes the high computational demands required to train the machine learning model. After several rounds of training, the AI<\/strong>\u200b\u200bmodel is able to predict interactions between a larger number of atoms over longer timescales.<\/p>\n\n\n\n “Springfield performed simulations for hundreds of nanoseconds on a massive system of thousands of atoms\u2014a task that would have been impossible using traditional methods. AI\/machine learning has truly transformed the landscape of complex materials research,” Lu said.<\/p>\n\n\n\n