Projects : Machine learning
Big-data collection
NLP based database generation from scientific document
As the interest in machine learning increases, the interest in databases for use in machine learning is also increasing. Despite the increasing demand for databases in the material field, there is no clear database to date. Therefore, in this project, we intend to build a database based on research results in the field of materials existing in Internet documents through natural language processing techniques.
Machine learning application
Assessing Additive Compositions for Glass Using MLP, MD, and DFT-based Simulations
Amorphous silicon dioxide (SiO₂) is a fundamental oxide material with extensive scientific and commercial applications, attributed to its excellent thermal properties, dielectric characteristics, and optical transmission performance. To enhance the transparency of silica glass—the amorphous form of SiO₂—dopants such as germanium (Ge) or titanium (Ti) are often introduced to increase optical light transmission. Using ab initio molecular dynamics, potential datasets are generated on Ge- and Ti-doped amorphous SiO₂ to train machine learning potentials. These trained models enable the reproduction of amorphous structures with quantum-level accuracy. The dielectric constant, and consequently the refractive index, of the Ge/Ti-doped amorphous SiO₂ is then calculated to assess the system’s transmission coefficient.
Full-Cycle AI Technology for Developing Ferroelectrics toward In-Memory Computing
In-memory computing (IMC) is a computing technology that performs data storage and computation within the system’s main memory, overcoming the von Neumann bottleneck. The IMC systems require superior memory devices with both scalable, non-volatile storage and low-power, high-endurance, high-speed computing properties. Ferroelectric RAM (FeRAM) is a promising device for IMC systems, which is non-volatile and low power operation, but its large cell-size and weak endurance are still in challenges. A full-cycle AI-based materials design integrates materials, process, microstructure, device, and circuit level properties through AI-driven generation, prediction, and optimization. Using an automated data crawling system based on natural language processing (NLP) with experimental validation process, this design system can rapidly predict optimal recipes for superior FeRAM devices, addressing the shortcomings in current IMC applications.
Understanding the domain walls movement of ferroelectric HfO2 with machine learning potential
HfO2-based FeRAM has been spotlighted as a key material for future semiconductor technologies such as logic-in-memory, neuromorphic computing, and negative capacitance FET, but it is facing difficulties in practical use due to its high coercive field and low domain wall velocity. Therefore, this study aims to understand and investigate the phenomenon related to the movement of domain walls under the electric field of HfO2-based ferroelectric materials at the atomic level through machine learning and molecular dynamics simulation.
Accelerated materials design of protonic ceramic electrolysis cell (PCEC) using first-principles calculation and machine learning
Protonic ceramic electrolysis cell is a next-generation hydrogen production system that can be operated at low temperatures and generates pure hydrogen from the fuel electrode. High chemical stability and proton conductivity material design is required as air electrode and electrolyte materials for PCEC. Using first-principle calculations, the mechanisms of proton conduction and water splitting reactions are identified, and doping materials are explored. Based on the first principle calculation data, PCEC materials are optimized in a wide composition space through a material performance prediction machine learning model.
Discovery of organic cathode material for high performance organic Li-ion battery
Inorganic cathode materials that are currently commercialized have environmental problems such as large amount of carbon emissions generated during the process and the use of expensive metals such as Co and Ni. Due to these problems, interest in organic-based batteries is increasing. However, in order to use organic-based cathode materials, low power, low energy density problems must be solved. Unlike inorganic materials, organic-based cathode materials have numerous possible combinations, so it is difficult to develop them through experiments. Therefore, we want to search for organic cathode materials with optimal performance through screening based on first-principle calculation and machine learning.
AI based supercritical materials research
As global issues emerge, exploration and development of resources and energy in extreme environments such as space, deep sea, and polar regions are actively progressing. Supercritical materials that greatly exceed the physical properties of existing materials are indispensable for stable work. However, its slow development speed and high probability of failure are blocked the progressing. Therefore, in this project, 7 promising supercritical materials that are used in the future extreme environment were selected, and an AI platform that integrates material design/manufacturing, synthesis/analysis, and evaluation is built to develop those materials in a relatively short time.Prediction of polymer properties using machine learning from the composition of various raw materials
Various properties of copolymer generally come from their constituent units (monomer). Several factors including type of units, composition of units, manufacturing conditions can affect properties of copolymer. Traditionally, to achieve desired properties, trial-and-error experiment method have been used but the cost is high because a lot of samples to be prepared to test. In this project, to address the limitations, we predict properties of copolymer through the machine learning models. Experimental dataset comprised of over 350 features is used for training data. Based on the models, we aim to find out suitable materials and their optimized compositions for desired products.
END.
Completed projects
Development of Novel Resistance Switching Materials by using First Principle-based Machine Learning Platform
AIM-HS(Ab Initio calculation based Machine learning High-throughput Screening) platform is Ab initio calculation high-throughput screening system based on multi-dimensional machine learning analysis. This project is purposed to detect hidden correlation between calculated properties and make criterion based on big data of Ab initio calculation. The final goal is material designing based on needs for material property.
Development of machine learning model for predicting corrosion of high corrosion-resistant alloy plated steel
Alloy plating is used to prevent corrosion of steel construction materials, but it has limited lifespan under harsh environments. Through a machine learning model developed from alloy plated steel corrosion images and experimental environment data, we predict lifespan of construction materials in real marine and soil environments.