Artificial Intelligence for Materials Science

The group is dedicated to harnessing the potential of artificial intelligence (AI) in the field of material science. Our main objective is to tackle critical challenges such as advanced materials design, experimental data analysis, text and data mining for information retrieval. To achieve these goals, we use a range of AI methods, including classical regression models, Bayesian optimization and deep learning techniques. Through the integration of diverse AI techniques, we aim to optimize material science problems with greater efficiency and accuracy. By leveraging the power of AI, we aim to make significant contributions to the field of material science and push the boundaries of what is possible.

Our group is working to developing active learning approaches that combine machine learning algorithms, theory and experiments in a closed loop to accelerate alloy design (see Fig. 1). Specifically, we begin with generative models based on autoencoder methods that can learn the distribution of alloy compositions in a low-dimensional space. This enables the generation of candidate alloy components that exhibit exceptional performance characteristics. Next, we integrate domain knowledge into the ensemble model, which is comprised of neural networks and decision trees, to enable the prediction of the mean and variance. Finally, we combine our predicted results with exploration-exploitation strategies to recommend compositions that will be subject to experimentation in order to update the database. This active learning process is an effective means of addressing the issue of small alloy databases and optimizing the alloy design process.

Figure 1. An active learning framework for the targeted composition design and discovery of HEAs, which combines machine learning models, density functional theory calculations, thermodynamic simulations and experimental feedback. — **Figure 1.** An active learning framework for the targeted composition design and discovery of HEAs, which combines machine learning models, density functional theory calculations, thermodynamic simulations and experimental feedback.

**Figure 1.** An active learning framework for the targeted composition design and discovery of HEAs, which combines machine learning models, density functional theory calculations, thermodynamic simulations and experimental feedback.

Another project we are working on is text data mining. The vast majority of scientific information is presented in written form, making it difficult to analyze using traditional statistical methods. This is particularly true in the materials research field, where the main source of machine-interpretable data has been limited to structured property databases, which represent only a small portion of the knowledge available in scientific literature. Our group is working to analyze material science texts, such as journal articles, using data mining and natural language processing to extract this knowledge and provide ideas for material design. By doing so, we hope to demonstrate the potential for collective extraction of information and relationships from the vast body of scientific literature, paving the way for a broader approach to mining scientific literature for materials science purposes.

Our group also focuses on utilizing AI to analyze experimental data points more efficiently. For example, atom probe tomography (APT) can generate millions of data points including atomic coordinates and species. In the past, statistical methods such as special distribution maps were commonly used to analyze such experimental results. We aim to introduce machine learning methods such as convolutional neural networks to help analyze these experimental data points. Our ultimate goal is to obtain valuable information about alloy composition and structural information in an efficient and repeatable manner.