Researchers in Korea have developed a technology that automatically identifies the necessary precursor materials to synthesize specific target materials.
A joint research team led by Senior Researcher Gyoung S. Na from the Korea Research Institute of Chemical Technology (KRICT) and Professor Chanyoung Park from the Korea Advanced Institute of Science and Technology (KAIST) has developed an AI-based retrosynthesis methodology that predicts the required precursor materials solely based on the chemical formula of the target material without expensive material descriptors and chemical analysis.
Precursor materials refer to all essential materials required in the synthesis process of the desired target material.
In recent years, materials discovery has become a crucial task across various industries, including batteries and semiconductors. Traditionally, finding the right intermediate materials for synthesis has required costly and repetitive experiments. However, there is a growing demand for utilizing AI to identify these materials efficiently.
Existing AI-based technologies for predicting material synthesis processes have primarily focused on organic materials, such as drug compounds, whereas research on inorganic materials has been relatively insufficient. This is because inorganic compounds, such as metals, possess complex structures and diverse elemental compositions, making it challenging to determine synthesis pathways.
The research team developed an innovative AI technology that can learn the inverse process of predicting the necessary precursor materials of the target material using only its chemical formula.
Previously, KRICT developed and transferred the 'ChemAI' platform in 2022, which allows users to predict synthesis information without advanced programming skills and expensive hardware infrastructures.
This newly developed technology overcomes the challenges posed by the complex 3D structures of inorganic materials, such as atomic arrangements and bonding information. Instead, the AI analyzes the types and ratios of elements present in the target material and calculates the thermodynamic formation energy differences to identify precursors that facilitate easier synthesis reactions.
To improve the accuracy of precursor material predictions, the team employed a deep neural network specialized in chemical data. The AI model was trained on approximately 20,000 published research papers detailing material synthesis processes and precursor materials.
The AI model was tested on around 2,800 synthesis experiments that were not provided in the training dataset. The evaluation results showed that it successfully predicted the necessary precursor materials in over 80% of cases within just 0.01 seconds by utilizing GPU acceleration.
Looking ahead, the research team plans to expand the training dataset through research projects of KRICT to achieve 90% prediction accuracy. By 2026, they aim to establish a web-based public service for AI-based materials discovery. Future research will focus on fully-automated materials discovery that predicts both precursor materials and synthesis pathways based solely on the target material's chemical formula.
The research team emphasized the novelty of their approach, stating, "Unlike conventional precursor prediction AI models that are limited to specific material types, our AI can predict precursor materials universally, regardless of the applications of the target materials."
KRICT President Young-Kuk Lee added, "This research is expected to enhance the efficiency of new material development across various industries."
