The system conducts large-scale simulation of 3D seismic wave propagation, and results are improved based on real-time data assimilation using observation and machine-learning.

Scope of use case

This system provides accurate information for evacuation in an earthquake

disaster.

Description

1 New direction in supercomputing:

The majority of SCD/ITC/U. Tokyos (Supercomputing

Research Division, The University of Tokyo) supercomputer

system users belong to the fields of computational science

and engineering, including engineering simulations (fluid

dynamics, structural dynamics, and electromagnetics), earth

sciences (atmosphere, ocean, solid earth, and earthquakes),

and material sciences. Recently, the number of users related

to data science, machine learning, and artificial intelligence

(AI) has been increasing. Examples of new research topics

are weather prediction by data assimilation, medical image

recognition, and human genome analyses. Towards Society

5.0, a new type of method for solving scientific problems that

integrates simulation (S), data (D) and learning (L)

(S+D+L) is emerging.

2 BDEC: Big data and extreme computing

The BDEC system (Big data and extreme computing), which

is scheduled to be introduced to SCD/ITC in April 2021, is a

hierarchical, hybrid, heterogeneous (h3) system. The BDEC

is the platform for integration of simulation, data and

learning (S+D+L), and consists of computing nodes for

computational science, those for data science/machine

learning, and those for integration. The aggregated peak

performance of the BDEC system is expected to be 40+

PFLOPS with aggregated memory bandwidth: 5,00+PB/sec.

It comprises three types of compute nodes: simulation

nodes (SIM, 90 % of total resources) for traditional

supercomputing applications; data/learning nodes (DL,

5 %) for data and learning; and integration nodes (INT,

5 %). The architecture of SIM and INT is necessary to be the

same, while that of DL can be different. Some of the DL nodes

would be connected to external resources (e.g. data storage,

servers, sensor networks, etc.) directly through an external

network (e.g., SINET, Japan). DL and INT would share a fast

file system (capacity: 4+PB, bandwidth: 2+TB/sec), while all

nodes share a large-scale file system (shared file system, 60

+ PB, 500 + GB/sec).

3 h3-Open-BDEC: Innovative software platform for

integration of (S+D+L)

We develop an innovative software platform h3-Open-

BDEC for integration of (S+D+L) and evaluate the effects of

integration of (S+D+L) on the BDEC (Figure.A.1). The h3-

Open-BDEC is designed for extracting the maximum

performance of the supercomputers with minimum energy

consumption focusing on (1) Innovative method for

numerical analysis with high-performance/high-

reliability/power-saving based on the new principle of

computing by adaptive precision, accuracy verification and

automatic tuning, and (2) Hierarchical data-driven approach

(hDDA) based on machine learning. This work would be

supported by the Japanese Government from FY.2019 to

FY.2023 (JSPS Grant-in-Aid for Scientific Research (S), P.I.:

Kengo Nakajima (ITC/U.Tokyo)).

Figure.A.1 Overview of h3-Open-BDEC

In the data driven approach (DDA), the technique of machine

learning is introduced for predicting the results of

simulations with different parameters. DDA generally

requires a lot of simulations for generation of teaching data.

We propose the hDDA, where simplified models for

generating teaching data are constructed automatically by

machine learning with feature detection, model order

reduction (MOR), uncertainty quantification (UQ), sparse

modelling and adaptive mesh refinement (AMR).

The h3-Open-BDEC is the first innovative software platform

to realize integration of (S+D+L) on supercomputers in the

Exascale Era, where computational scientists can achieve

such integration without support from other experts. Source

codes and documents are open to the public for various kinds

of computational environments. This integration by h3-

Open-BDEC enables significant reduction of computations

and power consumption compared to those by conventional

simulations.

The idea of h3-Open-BDEC is extension of that of ppOpen-

HPC [232] ppOpen-HPC is part of a (five + three)-year

project (FY.20112015, FY.2016-2018) supported by JST-

CREST and DFG-SPPEXA in Germany.

Possible applications on the BDEC system with h3-Open-

BDEC are combined simulations/data assimilations for

climate/weather simulations and earthquake simulations,

and real-time disaster simulations, such as flood, earthquake

and tsunami.

4 System for real-time earthquake simulation with data

assimilation

The system conducts large-scale simulation of 3D seismic

wave propagation, and results are improved based on real-

time data assimilation using observation and machine-

learning. Observations of seismic activities at more than 2

000 points in Japan are obtained by JDXnet developed by

ERI/U. Tokyo through SINET in real time. Construction of the

detailed and accurate underground model is crucial for

accurate simulations. An optimized underground model is

also constructed by integration of (S+D+L).

Interested in the same or similar project?
Submit a request and get a free project evaluation.