Registered Data

[02557] Collaboration of machine learning and physics-based simulation on earthquake disasters

  • Session Date & Time :
    • 02557 (1/2) : 1C (Aug.21, 13:20-15:00)
    • 02557 (2/2) : 1D (Aug.21, 15:30-17:10)
  • Type : Proposal of Minisymposium
  • Abstract : In Japan, strong-ground-motion data have been accumulated over a quarter of a century by a nationwide observation network. Large-scale physics-based simulations using HPC have enabled evaluations that consider the uncertainties of natural phenomena. Studies have begun to make up for the observation data related to large-scale disasters, which are currently lacking, using these simulation data. Furthermore, as studies toward hazard and risk assessment, surrogate modeling and damage assessment modeling by machine learning using observation and simulation data are being performed. In this mini-symposium, we will introduce the collaborative research of machine learning and physics-based simulation mainly on earthquake disasters.
  • Organizer(s) : Takahiro Maeda, Takuzo Yamashita, Asako Iwaki, Ryuta Imai
  • Classification : 86A15, 68T07
  • Speakers Info :
    • Shin Aoi (National Research Institute for Earth Science and Disaster Resilience)
    • Ryuta Imai (Mizuho Research & Technologies)
    • Takahiro Maeda (National Research Institute for Earth Science and Disaster Resilience)
    • Asako Iwaki (National Research Institute for Earth Science and Disaster Resilience)
    • Tomohisa Okazaki (Riken)
    • Hirotaka Hachiya (Wakayama University)
    • Takuzo Yamashita (National Research Institute for Earth Science and Disaster Resilience)
    • Bahareh Kalantar (Riken)
  • Talks in Minisymposium :
    • [03553] Optimal Transport in Seismic Wave Analysis
      • Author(s) :
        • Tomohisa Okazaki (RIKEN)
      • Abstract : An appropriate measure of the similarity between waveforms is crucial for seismic data analysis and modeling. The use of the Wasserstein distance in optimal transport theory has received attention in seismology because it captures time difference of waveforms. This presentation introduces two research directions: (1) converting acceleration envelopes from long to short periods for predicting ground motions caused by scenario earthquakes; (2) the sliced Wasserstein distance between seismograms to efficiently measure the similarity of oscillating seismic signals. These applications support the effectiveness of the Wasserstein distance as a similarity measure of seismic waveforms.
    • [03695] A smoothing scheme for seismic wave propagation simulation with SDWave
      • Author(s) :
        • Ryuta Imai (Mizuho Research & Technologies, Ltd.)
      • Abstract : We propose a smoothing scheme SDWave for seismic wave propagation simulation. The SDWave is based on a diffusionized wave equation with the fourth-order spatial derivative term. We mathematically explain some properties of the equation and how the SDWave works for smoothing. Then we give two discretization methods, FDM and mixed FEM, of the SDWave and apply it to the wave equation. This numerical experiment reveals that the SDWave is effective for filtering short wavelength components.
    • [03740] Automated Building Damage Assessment using Multi-scale Siamese Deep Learning Network
      • Author(s) :
        • Bahareh Kalantar (RIKEN AIP)
        • Naonori Ueda (RIKEN AIP)
      • Abstract : Timely information on building damage location is vital for emergency responders after natural disasters. Our proposed Multi-scale Siamese Building Damage Assessment model assesses damage by localizing buildings and classifying damage level into four categories. The model employs a multi-scale block to handle buildings of varying sizes. The results indicate the model's effectiveness, although it struggles with classifying minor and major damage.
    • [03984] Position-dependent inpainting for ground motion interpolation
      • Author(s) :
        • Hirotaka Hachiya (Wakayama University)
      • Abstract : Acquiring continuous spatial data is essential to assess the damaged area just after the earthquake. To this purpose, we propose a framework of supervised spatial interpolation and apply highly advanced deep inpainting methods with the introduction of position-dependent partial convolution, where convolution kernel weights are adjusted depending on their position on an image based on the trainable position-feature map. We show the effectiveness of our proposed method, through experiments using ground-motion data.
    • [04162] A quarter century of data from K-NET and KiK-net
      • Author(s) :
        • Shin Aoi (National Research Institute for Earth Science and Disaster Resilience)
        • Takashi Kunugi (National Research Institute for Earth Science and Disaster Resilience)
        • Wataru Suzuki (National Research Institute for Earth Science and Disaster Resilience)
        • Hiroyuki Fujiwara (National Research Institute for Earth Science and Disaster Resilience)
      • Abstract : Based on the lessons learned from the 1995 Kobe earthquake, the National Research Institute for Earth Science and Disaster Resilience (NIED) has constructed K-NET and KiK-net, nationwide strong-motion observation networks that homogeneously cover the entire country. The strong-motion database obtained from these world's largest strong-motion observation networks contains nearly one million archived records. In this presentation, these observation networks and databases will be introduced and the utilization of the data will be discussed.
    • [04544] Physics-based long-period ground motion simulation for megaquakes
      • Author(s) :
        • Takahiro Maeda (National Research Institute for Earth Science and Disaster Resilience)
      • Abstract : There are limited seismic observation records directly linked to damage, such as ground motions caused by huge earthquakes and those near seismic faults. In order to evaluate such ground motions, physics-based seismic-ground-motion simulation using the three-dimensional subsurface structure and seismic-source models is carried out, which are used to clarify the causes of damage and predict future seismic motions. In this presentation, we will introduce ground-motion simulation methods and examples of their application to huge earthquakes.
    • [04967] Linkage of physics simulation and machine learning towards seismic risk assessment
      • Author(s) :
        • Takuzo Yamashita (NIED)
        • Jun Fujiwara (NIED)
        • Asako Iwaki (NIED)
        • Hiroyuki Fujiwara (NIED)
      • Abstract : The authors are developing a seismic risk assessment method with physical simulation and machine learning. Response surfaces of seismic demand are modeled by Gaussian process regression. Low-dimensional features of seismic motions with auto-encoder were used as input data. An active learning method using Bayesian optimization was developed to construct the model with a small number of samplings. As a result, proposed model using samples less than 1/10th of the total data successfully predicted correct values.
    • [05004] Construction of strong motion database for data-driven ground-motion prediction models
      • Author(s) :
        • Asako Iwaki (National Resesarch Institute for Earth Science and Disaster Resilience)
        • Nobuyuki Morikawa (National Resesarch Institute for Earth Science and Disaster Resilience)
        • Takahiro Maeda (National Research Institute for Earth Science and Disaster Resilience)
        • Hiroyuki Fujiwara (National Resesarch Institute for Earth Science and Disaster Resilience)
      • Abstract : We have been developing a strong-motion observation database as an infrastructural database utilized for seismic hazard assessment, from which data-driven regression models for ground-motion prediction (ground-motion models; GMMs) are to be constructed. The database is “biased” because there are insufficient number of records with large magnitudes and short distances. Consequently, GMMs are incapable of predicting such ground motion. To overcome this issue, we attempt to utilize simulated ground motion data to supplement the observation database.