AI-guided experiments speed scientific discovery

Finding a needle in a haystack is the quintessentially impossible task. But what if new tools could make it straightforwardly achievable? Imagine if, instead of searching through everything by hand, you could portion out small piles of hay and use magnets. Artificial intelligence (AI) can act as a magnet for scientific solutions, pulling important information from a mountain of possibilities. 

But AI can only do so much. If the proverbial haystack is too big, even the most powerful system can be stymied. Effectively using AI requires intelligently involving domain science expertise in the process. A team of scientists brought AI, high-throughput experimentation, and materials science knowledge together to speed up the discovery process.

The combination worked. The team, led by researchers from Pacific Northwest National Laboratory (PNNL) and Argonne National Laboratory, identified combinations of solvents that can dissolve three times more of a compound proposed as part of an energy-efficient redox flow battery. They succeeded by quickly narrowing their search to less than 10 percent of possible combinations. The findings appear in Nature Communications.

The team included experts from complementary backgrounds, all focused on creating a platform that can intelligently perform high-throughput experiments. They explored a range of organic solvent mixtures to design an optimal electrolyte system for redoxmer-based flow batteries. 

“Often, people look at an automated system as a way to speed up discovery by dramatically increasing the number of experiments that can be done,” said Vijay Murugesan, a PNNL materials scientist and co-corresponding author on the paper. “We wanted to speed up discovery with increased efficiency using AI for science.”

While the platform specifically targets electrolyte mixtures for energy storage, the general process can be applied to other systems. This could be most useful for problems with a vast array of potential solutions within a constrained system, the researchers said.

High-throughput data for artificial intelligence

Rather than running experiments independently, the high-throughput experimentation team gathered data to fill in gaps for the AI team’s algorithm. Often, the type of data the AI model needs is not available for laboratory systems. The algorithm then has to be trained on computational results, which can lead to additional biases.

On the experimental side, determining optimized solvent mixtures is a massive problem. “We identified 2,000 possible combinations,” said Yangang Liang, a co-corresponding author and expert in high-throughput experimentation at PNNL. “That is an impractical number of combinations to test even with our robotic system. While the robot can do experiments faster, it still requires chemicals and energy.”

Identifying the most promising options without AI would still have required hundreds of experiments. To narrow their search, the team targeted their initial data collection based on known gaps in the training sets for the AI model. Feeding the high-fidelity experimental data into the model led to a better-trained system, which in turn gave better predictions for the next round of experiments.

“Our approach is incredibly efficient,” said Murugesan. “We’re leveraging the speed of high-throughput and human intuition to better train AI.”

The power of collaborative data

The product of this collaboration is twofold: first is identifying the solvent mixture, the scientific goal of the work. The second is creating a high-fidelity dataset from experimental data. The team hopes that others will be able to make use of the data for future work beyond exploring solvent mixtures for organic redox flow batteries.

“We were intentional in our approach to creating high-fidelity data that can help build better predictive models,” said Murugesan. “Our process was informed by the broad expertise of our team, something made possible by the Department of Energy’s investment in center-scale work. Centers specialize in these types of ambitious ideas that require multiple disciplines to come together.”

The project was funded through the Joint Center for Energy Storage Research (JCESR), an energy storage research effort that brought together six national laboratories and 10 universities from 2018-2023. 

“This work was really inspired by the late George Crabtree, the founding director of JCESR,” said Murugesan. “We went to him with the idea to use PNNL’s high-throughput capability for electrolyte discovery, but he challenged us to think bigger and collaborate with the AI team. Through his inspiration, we learned that together we can produce impactful results faster by integrating AI models and robotic platforms.” 

Taking steps to a self-driving lab 

The materials-informed data produced by the team is the type necessary for creating the effective AI systems that will drive the experimental loops in autonomous lab spaces. “I see these types of workflows as central to a new paradigm in materials discovery,” said Hieu Doan, a co-corresponding author who led the AI work. 

“I’m excited to see the future of collaboration between AI researchers and materials scientists,” added Karl Mueller, a co-author of the paper and the Director of the Program Development Office for the Physical and Computational Sciences Directorate. “Accelerating materials discovery is critical to solving energy storage problems.”

In addition to Liang, Murugesan, and Mueller, Juran Noh and Heather Job contributed to the project from PNNL. The Argonne team included Doan, Lily Robertson, Lu Zhang, and Rajeev Assary. Many of the collaborators on this work are part of the newly launched Energy Storage Research Alliance Energy Innovation Hub.