Emergence of crack tip plasticity in semi-brittle -Fe

Suzudo, Tomoaki; Ebihara, Kenichi; Tsuru, Tomohito; Mori, Hideki*

Journal of Applied Physics, 135(7), p.075102_1 - 075102_7, 2024/02

https://doi.org/10.1063/5.0178940

Fracture of body centred cubic (bcc) metals and alloys below the ductile-to-brittle transition temperature is brittle. This is theoretically explained by the notion that the critical stress intensity factor of a given crack front for brittle fracture is smaller than that for plasticdeformation; hence, brittle fracture is chosen over plastic deformation. Although this view is true from a macroscopic point of view, such brittle fracture is always accompanied by small-scale plastic deformation in the vicinity of the crack tip, i.e. crack tip plasticity. This short paper investigates the origin of this plasticity using atomistic modeling with a recently developed machine-learning interatomic potential of -Fe. The computational results identified the precursor of crack tip plasticity, i.e. the group of activated atoms dynamically nucleated by fast crack propagation.

Machine learning molecular dynamics reveals the structural origin of the first sharp diffraction peak in high-density silica glasses

Kobayashi, Keita; Okumura, Masahiko; Nakamura, Hiroki; Itakura, Mitsuhiro; Machida, Masahiko; Urata, Shingo*; Suzuya, Kentaro

Scientific Reports (Internet), 13, p.18721_1 - 18721_12, 2023/11

https://doi.org/10.1038/s41598-023-44732-0

The first sharp peak diffraction peak (FSDP) in the structure factor of amorphous materials is thought to reflect the medium-range order structure in amorphous materials, and the structural origin of the FSDP has been a subject of ongoing debate. In this study, we employed machine learning molecular dynamics (MLMD) with nearly first-principles calculation accuracy to investigate the structural origin of the FSDP in high-density silica glass. First, we successfully reproduced various experimental data of high-density silica glass using MLMD. Furthermore, we revealed that the development (or reduction) of the FSDP in high-density silica glass is characterized by the deformation behavior of ring structures in Si-O covalent bond networks under compression.

Machine learning molecular dynamics simulations for evaluation of high-temperature properties of nuclear fuel materials

Kobayashi, Keita; Nakamura, Hiroki; Itakura, Mitsuhiro; Machida, Masahiko; Okumura, Masahiko

Materia, 62(3), p.175 - 181, 2023/03

https://doi.org/10.2320/materia.62.175

no abstracts in English

Optimization of mercury flow with microbubbles in the target-vessel design by means of machine learning

Kogawa, Hiroyuki; Futakawa, Masatoshi; Haga, Katsuhiro; Tsuzuki, Takayuki*; Murai, Tetsuro*

JAEA-Technology 2022-023, 128 Pages, 2022/11

JAEA-Technology-2022-023.pdf:9.0MB

In a mercury target of the J-PARC (Japan Proton Accelerator Research Complex), pulsed proton beams repeatedly bombard the flowing mercury which is confined in a stainless-steel vessel (target vessel). Cavitation damage caused by the propagation of the pressure waves is a factor of the life of the target vessel. As a measure to reduce damages, we developed a bubbler to inject the gas microbubbles into the flowing mercury, which can reduce the pressure waves. To operate the mercury target vessel stably with the 1 MW high-intensity proton beams, further reduction of the damage is required. The bubbler setting position should be closer to the beam window to increase the bubble population, which could enhance the reduction effect on the pressure waves and damage. However, the space at the beam window of the target vessel is restricted. The bubbler design and setting position as well as the vane design for the mercury flowing pattern are optimized by means of a machine learning technique to get more suitable bubble distribution, increasing in bubble population and optimizing bubble size nearby the beam window of the target vessel. The results of CFD analyses performed with 1000 cases were used for machine learning. Since the flow rate of mercury affects the temperature of the target vessel, this was used for the constraint condition. As a result, we found a design of mercury target vessel that can increase the bubble population by ca. 20% higher than the current design.

Machine learning molecular dynamics simulations toward exploration of high-temperature properties of nuclear fuel materials; Case study of thorium dioxide

Kobayashi, Keita; Okumura, Masahiko; Nakamura, Hiroki; Itakura, Mitsuhiro; Machida, Masahiko; Cooper, M. W. D.*

Scientific Reports (Internet), 12(1), p.9808_1 - 9808_11, 2022/06

https://doi.org/10.1038/s41598-022-13869-9

no abstracts in English

G-HyND: A Hybrid nuclear data estimator with Gaussian processes

Iwamoto, Hiroki; Iwamoto, Osamu; Kunieda, Satoshi

Journal of Nuclear Science and Technology, 59(3), p.334 - 344, 2022/03

https://doi.org/10.1080/00223131.2021.1971120

A hybrid nuclear data estimator (G-HyND) based on a machine learning technique with Gaussian processes (GP) was developed. G-HyND estimates cross-sections from a hybrid training dataset composed of an experimental dataset and an analytical dataset based on a nuclear physics model, and generates the cross-section datasets including the dataset's uncertainty information. It was demonstrated that an experimental dataset and a physics model-based analytical dataset perform a complementary role in nuclear data generation, and that the generated nuclear data from the hybrid training dataset are more reasonable than only those from the experimental dataset. Furthermore, solutions for two inherent GP problems, i.e., overfitting and computational cost, are presented within the G-HyND framework.

Atomic Energy Society of Japan 2021 Annual Meeting, Joint session of "Research Committee for Nuclear Data" and "Subcommittee on Nuclear Data"; Activity report of research committee for nuclear data in the fiscal years of 2019 and 2020, 4; Trends of major evaluated nuclear data file in the world

Suyama, Kenya

Kaku Deta Nyusu (Internet), (130), p.29 - 34, 2021/10

http://www.aesj.or.jp/~ndd/ndnews/pdf130/No130-07.pdf

This manuscript describers the appearance of Japanese Evaluated Nuclear Data Library (JENDL) for Europe, the status of the main nuclear data library of European countries, i.e., Joint Evaluated Fission and Fusion (JEFF) Nuclear Data Library and the future of evaluation of the nuclear data, based on the experience of working at OECD/NEA Data Bank which manages the development of JEFF.

Session of Sub-committee on Nuclear Data; Application of advanced data science to nuclear data, 3; Application of machine learning to nuclear-data evaluation study and high-energy particle transport calculation

Iwamoto, Hiroki

Kaku Deta Nyusu (Internet), (129), p.19 - 25, 2021/06

http://www.aesj.or.jp/~ndd/ndnews/pdf129/No129-03.pdf

no abstracts in English

Identification of advanced spin-driven thermoelectric materials via interpretable machine learning

Iwasaki, Yuma*; Sawada, Ryoto*; Stanev, V.*; Ishida, Masahiko*; Kirihara, Akihiro*; Omori, Yasutomo*; Someya, Hiroko*; Takeuchi, Ichiro*; Saito, Eiji; Yorozu, Shinichi*

npj Computational Materials (Internet), 5, p.103_1 - 103_6, 2019/10

https://doi.org/10.1038/s41524-019-0241-9

Studies on learning by detectiong impasse and by resuling it for building large scale knowledge base for autonomous plant

*

PNC TJ9604 98-001, 83 Pages, 1997/03

PNC-TJ9604-98-001.pdf:4.54MB

Recently, due to the tremendous improvement of information infrastructures such as networking facilities, the idea of a large scale knowledge base with realtime operations for technological plants has emerged. The major bottleneck for building a large scale knowledge base for an autonomous plant lies in its design phase. The acquisition of knowledge from human experts in an exhaustive way is extremely difficult, and even if it were possible, the maintenance of such a large knowledge base for realtime operation is not an easy task. The autonomous system having just incomplete knowledge would face with so many problems that contradicts with the system's current beliefs and/or are novel or unknown to the system. Experienced humans can manage to do with such novelty due to their generalizing ability and analogical inference based on the repertoire of precedents, even if they with new problems. Moreover, through experiencing such breakdowns and impasse, they can acquire some novel knowledge by their proactive attempts to interpret a provided problem as well as by updating their beliefs and contents and organization of their prior knowledge. We call such a style of learning as impasse-driven learning, meaning that learning dose occur being motivated by facing with contradiction and impasse. The related studies concerning with such a style of learning have been studied within a field of machine learning of artificial intelligence so far as well as within a cognitive science field. In this paper, we at first summarize an outline of machine learning methodologies, and then, we detail about the impasse-driven learning. We discuss that from two different perspectives of learning, one is from deductive and analogical learning and the other one is from inductive conceptual learning (i.e., concept formation or generalization-based memory). The former mainly discuss about how the learning system updates its prior beiiefs and knowledge so that it can explain away the ...