Domestic researchers have developed technology that automatically informs what final material substances are needed to create specific materials.
On the 12th, Senior Researcher Na Kyung-seok of the Korea Institute of Chemical Technology and Professor Park Chan-young of the Korea Advanced Institute of Science and Technology (KAIST) announced that they have developed an artificial intelligence (AI) methodology to predict the raw materials (precursor materials) necessary for synthesis using only the chemical formula information of the target substance. The research results were presented at the Conference on Neural Information Processing Systems (NeurIPS) in December.
Advanced materials are very important in various industrial sectors such as batteries and semiconductors. To synthesize the desired material, intermediate substances must first be identified, and there is a high demand to find them using artificial intelligence (AI) instead of costly and repetitive experiments. However, existing AI-based technologies have focused on organic materials like new drugs, with relatively less research on inorganic materials. This is because synthesizing inorganic compounds like metals is challenging due to their complex structures and diverse elements.
The research team developed a new AI method that predicts the precursor materials A, B, and C needed to create a target substance X, using only the chemical formula of X. Previously, the Korea Institute of Chemical Technology developed the 'ChemAI' platform, which predicts the information required for material synthesis without the need for complex coding or server setup, and transferred the technology in 2022.
Unlike existing prediction technologies, this new technology does not require the complex three-dimensional structure of inorganic materials, such as atomic structures or bonding information. Instead, it examines the types and proportions of included elements. It calculates the thermodynamic formation energy differences between these elements and the target substance to identify precursors that facilitate the synthesis reaction.
To improve the accuracy of precursor material predictions, the research team configured a deep artificial neural network specialized in chemical data. The deep artificial neural network learned all information regarding material synthesis processes and precursor materials reported in approximately 20,000 papers. As a result of predicting the required precursor materials for approximately 2,800 material synthesis experiments that were not previously shown in the AI training process, it achieved success in more than 8 out of 10 attempts. Through graphics processing unit (GPU) acceleration, the prediction time was reduced to about 0.01 seconds.
The research team noted, "Unlike existing precursor material prediction AIs that could only be applied to specific types of substances, this research allows for the universal prediction of precursor materials regardless of the type of target substance." Lee Young-guk, director of the Korea Institute of Chemical Technology, expressed, "I expect this will contribute to improving research efficiency in various industries needing new material development."
The research team plans to expand its dataset through research projects at the Korea Institute of Chemical Technology, improving precursor material prediction accuracy to over 90%, and to establish a web-based public service around 2026. Additionally, through further research, they aim to provide complete predictions of not only precursor materials but also the material synthesis process using the chemical formula of the target substance, achieving "fully automated AI-based materials retrosynthesis."
Reference materials
arXiv(2024), DOI: https://doi.org/10.48550/arXiv.2410.21341