Fellowship - Topic 6: Super Accelerating Scientific Simulation with Physical Model Embedded Machine Learning Method for Real-Time Digital-Twin in Cyber Physical Systems
Japan

Odsylacz dla stanowiska: 000138

Lokalizacja: Japan

Oferta wazna do: 15/03/2021

Data publikacji oferty: 09/12/2020

Wynagrodzenie: C10M JPY

Typ zatrudnienia: Toshiba Fellowship

Rodzaj dzialalnosci: Research & Development

Topic 6: Super accelerating scientific simulation with physical model embedded machine learning method for real-time digital-twin in cyber physical systems

Background

Real-time digital-twin concepts as an inter-mediator taking care of the interaction among human, nature and infrastructure systems would be to simulate the complex real world interactions as well as to run the real world in the cyber physical systems [1][2].
Digital-twin computing could well improve infrastructure resilience for various risks, and to optimize design, operation and maintenance in cyber physical systems.
Scientific simulations such as fluid mechanics, non-linear mechanics, multi-physics and multi-scale phenomena are implementations of theoretical models, they may run very slowly, even on modern computer hardware; for example, global climate models may take thousands of CPU-hours.
It is necessary to highly accelerate simulations --computer implementations of mathematical modelsto test out new concepts and ideas, and explore potential behaviors of infrastructure systems and natural systems [3][4].

Toshiba has owned the data, model and know-how of infrastructure systems, and has potential and environment to develop digital-twin computing technologies [1][2].

Suggested Research Proposals

  • Accelerating finite-difference or finite-element simulation on Eulerian frame with machine learning method such as Lagrangian neural networks, Hamiltonian neural networks, deep neural networks, and convolutional networks [5][6].
  • Physical modeling method embedded machine learning for continuum mechanics or non-liner dynamical simulation based on physical law exploration or reduced order modeling method such as proper orthogonal decomposition, Galerkin projection, and sparse identification of nonlinear dynamical systems [7].
  • Accelerating hybrid simulations of multi-agent simulation and physical simulation for a large number of risk scenarios in infrastructure systems [8].
  • Making simulation based high-dimensional data predictable method such as randomly distributed embedding [9].

Related Papers

[1] https://www.toshiba.co.jp/iot/index_en.htm
[2] https://www.toshiba-energy.com/en/transmission/product/iot.htm
[3] https://www.infoq.com/news/2020/03/deep-learning-simulation/
[4] https://arxiv.org/abs/2001.08055
[5] https://arxiv.org/abs/2003.04630
[6] https://arxiv.org/abs/1906.01563
[7] https://arxiv.org/abs/1701.03424
[8] https://arxiv.org/abs/2004.05185
[9] https://www.pnas.org/content/115/43/E9994


 

Wstecz