Princeton researchers have developed an artificial intelligence (AI) tool that uses a large language model to predict the behavior of crystalline materials. This development has significant implications for advancing technologies such as batteries and semiconductors. While computer simulations are commonly used in crystal design, this new method takes advantage of a large language model, similar to those used in text generators like ChatGPT.
By analyzing text descriptions of crystal properties, including atomic bond lengths and angles, as well as electronic and optical measurements, the AI tool can predict the properties of new materials with greater accuracy and comprehensiveness compared to existing simulations. Ultimately, this could expedite the process of designing and testing new technologies.
The researchers constructed a text benchmark consisting of over 140,000 crystal descriptions obtained from the Materials Project. Using this benchmark, they trained an adapted version of a large language model called T5, originally developed by Google Research. The tool’s predictive capabilities were tested on known crystal structures, ranging from table salt to silicon semiconductors. With proven efficacy, the researchers now plan to apply the tool to the discovery of novel crystal materials.
Presented at the Materials Research Society’s Fall Meeting in Boston on November 29th, this method serves as a new benchmark that can potentially hasten materials discovery for various applications, according to Adji Bousso Dieng, senior study author and assistant professor of computer science at Princeton.
The research paper, titled “LLM-Prop: Predicting Physical And Electronic Properties Of Crystalline Solids From Their Text Descriptions,” has been published on the arXiv preprint server.
Current AI-based tools for crystal property prediction rely on graph neural networks, but these have computational limitations and struggle to accurately represent the intricacies of atomic geometry, bond lengths, and the resulting electronic and optical properties. Dieng’s team is the first to tackle this challenge using large language models.
Dieng explains that while significant progress has been made in computer vision and natural language processing, graph-based AI still lags behind. To overcome this limitation, the researchers leverage the power of large language models by translating the graph-based problem into a text format.
Craig Arnold, co-author of the study and Princeton’s Susan Dod Brown Professor of Mechanical and Aerospace Engineering, and vice dean for innovation, underscores the revolutionary approach of the language model-based method. He emphasizes the importance of tapping into humanity’s vast knowledge and effectively processing it to drive progress. This approach stands in contrast to current methods and represents a significant breakthrough in materials design.
By harnessing the capabilities of AI language models, researchers are poised to accelerate the discovery process of materials, unlocking new possibilities for technological advancements in various industries.
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it