Abstract : Artificial Intelligence has led to a paradigm shift in investigation in Materials Chemistry, with Machine Learning allowing informatics-based systematic calculations, predictions and discovery based on material databases pushing beyond the intrinsic limitations of first-principles calculations. The successful application requires development of novel methodologies inspired by the frontends of materials development in close synergy between Mathematics and Information Technology, areas where interdisciplinary collaborations are crucial and yet to date in their early phases.
[01431] Discovery of New Materials by Quantum Calculations and Artificial Intelligence
Format : Talk at Waseda University
Author(s) :
David Samuel Rivera Rocabado (Hiroshima University)
Mika Aizawa (Hiroshima University)
Takayoshi Ishimoto (Hiroshima University)
Abstract : Integrating artificial intelligence into real system first-principles calculations has the potential to transform the field of materials science and, ultimately, the world in which we live. In this presentation, the modeling and the prediction of the CO adsorption/activation on Ru nanoparticles exemplify that our new approach can be universally applied to predict the catalytic properties of any existing material and be used for the discovery of more functionalized materials.
[05452] Causal analysis of materials functionality by combining topological data analysis and physical model
Format : Talk at Waseda University
Author(s) :
Masato Kotsugi (Tokyo University of Science)
Abstract : Microscopic image data is key to developing low-power, high-speed electronic devices. However, the complex interactions in nanoscale magnetic materials are difficult to understand. We developed a new functional design theory called “extended Landau free energy model” that combines persistent homology and machine learning with free energy to automate the interpretation of the microscopic image. This model illustrates the physical mechanism and critical location of magnetization reversal and proposes a device structure with low energy consumption.
[05575] Machine learning for materials chemistry and chemical biology
Format : Talk at Waseda University
Author(s) :
Daniel Packwood (Kyoto University)
Abstract : This talk will review three recent success stories involving machine learning in materials science and chemistry: (1) the simulation of on-surface molecular self-assembly processes using machine-learned potentials and a novel genetic algorithm; (2) the design of a new organic semiconducting material with targeted band gap by a combination of unsupervised and supervised learning; (3) the prediction and verification of a new chemical compound for inducing cardiac tissue differentiation of stem cells.
[01449] Topological descriptor of thermal conductivity in amorphous Si
Format : Online Talk on Zoom
Author(s) :
Emi Minamitani (Osaka University)
Takuma Shiga (AIST)
Makoto Kashiwagi (Aoyama Gakuin University)
Ippei Obayashi (Okayama University)
Abstract : In this talk, we analyze the relationship between the atomic configuration of amorphous Si and the thermal conductivity of the material. A topological descriptor constructed by persistent homology successfully predicts the thermal conductivity using machine learning, and from the machine learning model, we can extract medium-range order structures related to thermal conductivity using inverse analysis of persistent homology. https://doi.org/10.1063/5.0093441
[05561] Fully automatized optimization of ring-opening reactions in lactone derivatives via 2-step machine learning
Format : Talk at Waseda University
Author(s) :
Aleksandar Staykov (Kyushu University)
Pierluigi Cesana (kyushu university)
Abstract : Cyclization and cycloreversion of organic compounds are fundamental kinetic processes in the design of functional molecules, molecular machines, and nano-switches. We present a fully automatic computational platform for the design of a class of 5- and 6- membered ring lactones by optimizing the ring-opening reaction rate. Starting from a minimal initial parent set, our program generates iteratively cascades of pools of candidate lactone derivatives where optimization and down-selection are performed not requiring human supervision at any stage. We use Density Functional Theory combined with transition state theory to elucidate the exact mechanism leading to the lactone ring opening. Based on the analysis of the reaction pathway and the frontier molecular orbitals, we identify a simple descriptor that can easily correlate with the reaction rate. The program is successful in identifying a large class of lactone derivatives with enhanced ring-opening properties. Our platform is modular and our current implementation for lactone could be further generalized to more complex systems via substitution of the quantum chemical and fingerprinting modules.
[01724] A Trial for the Realization of Material Pattern Informatics Using Interpretable AI
Format : Talk at Waseda University
Author(s) :
Yoh-ichi Mototake (Hitotsubashi University)
Abstract : It has been recently reported that highly accurate classification, regression, and generation can be achieved by interpolative modeling of complex scientific data using machine learning models with high expressive power, such as deep neural networks (DNNs). However, many of the machine learning models used there are nonlinear functions with a large number of parameters, making the interpretation of the training results very difficult. In thermodynamics, Gibbs extended the theory of thermodynamics, which was the theory of heat engines, to chemical reactions, which was a great development in science. This shows that science has been greatly advanced by the scientific insight of human beings, who derive general principles beyond mere interpolation models and boldly extrapolate them. On the other hand, it is sometimes difficult to apply such insights to systems with complex non-periodic structures, such as those found in nonlinear and nonequilibrium phenomena. To address this situation, we believe that it is important to collaborate between machine learning, which is good at building interpolation models for complex data, and humans, who can make bold extrapolations based on scientific insights, and are developing methods for interpreting machine learning training results to bridge the gap between the two.
In this presentation, we will discuss our recent research on machine learning frameworks that collaborate with scientists trying to reveal complex pattern dynamics in materials and their applications.
[05582] Density functional theory analysis for hydrogen sulfide removers with graphene
Format : Talk at Waseda University
Author(s) :
Takaya Fujisaki (Shimane University)
Abstract : Methane, attracting attention as a hydrogen carrier for solid oxide fuel cells, can be generated by methane fermentation using biomass, however, it is known to contain some amount of hydrogen sulfide. Since hydrogen sulfide reduces the power generation efficiency of fuel cells, it is desirable to remove as much hydrogen sulfide as possible. In this study, we used density functional theory to understand the hydrogen sulfide removing agent with graphene structure.
[01335] Understanding the role of defects and disorder in polycrystalline materials
Format : Talk at Waseda University
Author(s) :
Kulbir K Ghuman (Institut national de la recherche scientifique)
Abstract : The functionality of the materials used for energy applications is critically determined by the physical properties of small active regions such as dopants, dislocations, interfaces, grain boundaries, etc. The capability to manipulate and utilize the inevitable disorder in materials, whether due to the finite-dimensional defects (such as vacancies, dopants, grain boundaries) or due to the complete atomic randomness (as in amorphous materials), can bring innovation in designing energy materials. With the increase in computational material science capabilities, it is now possible to understand the complexity present in materials due to various defects resulting in pathways required for optimizing their efficiencies. In this talk, I will provide a critical overview of such computational advancements specifically for designing realistic materials with various types of defects. I will discuss the traditional approaches (implemented via tools such as density functional theory, and molecular dynamics) as well as modern approaches such as machine learning that exist for understanding the impact of defects and disorder present in polycrystalline materials, thereby identifying future opportunities for energy materials design and discovery.
