New polymer electrolytes for lithium-ion batteries can be developed using automated molecular design through machine learning.
In our efforts to answer contemporary scientific problems, machine learning is becoming increasingly pervasive. From the folding of proteins to image editing, machine learning is a well-established method for the discovery of new information.
Battery technology would benefit from the discovery of new materials. The application of machine learning to predict the performance of rechargeable batteries was limited by the quantity of the data set and the unsupervised nature of the model. And further obstacles remain. Researchers at Waseda University, Tokyo, in partnership with Fujitsu, have used machine learning to locate potential polymer materials for Li+ batteries far more effectively.
According to Kan Hatakeyama-Sato, a researcher at Waseda University’s Department of Applied Chemistry and the paper’s first author, AI can anticipate novel material structures with desirable properties. “However, the majority of candidate structures are recognized to be useless. They do not meet practical application characteristics such as synthetic simplicity, stability, and processability.”
The authors claim that utilizing AI alone to evaluate prospective materials is challenging because AI lacks the tacit knowledge that professionals have regarding the parameters that are most desirable for the material.
AI with professional guidance
Hatakeyama-Sato and the team examined the constraints of automating the discovery of new battery materials and considered how they could be circumvented with a little bit of expert assistance.
“Instead of teaching each exploratory criterion of materials to AI,” Hatakeyama-Sato explained, “the system was educated to learn the appearance of practical items.” Existing materials’ molecular structures were used to teach AI their fundamental characteristics (unsupervised machine learning). Our approach, which was inspired by existing species, could produce unheard-of material structures with greater performance. This technique is comparable to the design of new materials by researchers with extensive knowledge of existing materials.
Once the machine has generated a list of material possibilities, they were investigated using a new piece of hardware known as a digital annealer, a sophisticated computing device that permits researchers to efficiently investigate the candidate materials for their potential as battery materials.
“The AI system designed a new ion-conducting polymer,” Hatakeyama-Sato claimed. In addition to offering strong conductivity, the material met the criteria for solid-state electrolytes such as mechanical resilience, sufficient solubility, chemical stability and thermal stability.
Future obstacles confronting machine learning
Materials science might considerably profit from more efficient and advanced automation of the structural design process, particularly for organic materials. The researchers believe that this system will help us get closer to our goal. As a potential next step, AI may inform us not just of the optimal end-product materials, but also of the synthetic paths required to produce them.
Hatakeyama-Sato concluded, “Organic materials are difficult to design computationally due to their complicated architecture.” “However, as demonstrated by these results, new algorithms and hardware are gradually paving the way for the fully automated design of materials.”