Scholarlink

<p>Most existing Ground Motion Models (GMMs) are developed by specifying empirical functional forms derived from expert judgment, with regression techniques used to estimate model coefficients from recorded ground motion data. However, due to the complex and nonlinear nature of earthquake source and path effects, it is difficult to establish a direct analytical relationship between ground motion intensity measures and seismological predictors. This, combined with inherent natural variability, poses significant challenges for accurately characterizing ground motion using predefined functional forms. The emergence of deep learning methods offers a promising avenue for uncovering non-linear correlations among high-dimensional variables. In this study, the deep learning technique, Conditional Generative Adversarial Network (CGAN), is proposed as a novel, data-driven GMM for predicting horizontal-component spectral accelerations over a period range of 0 to 10&#xa0;s. The model is trained and evaluated using 11675 sets of recorded ground motions from the Engineering Strong motion database (ESM2.0). Model performance is rigorously evaluated through multiple metrics, including residual analysis and comparison with benchmark empirical GMMs. Results demonstrate that the CGAN outperforms traditional models in capturing complex spectral patterns and exhibits superior generalization with reduced prediction bias. Furthermore, a comparative analysis with other deterministic and probabilistic machine learning models developed using the same dataset highlights similarity in aleatory uncertainty but notable differences in epistemic uncertainty estimation, attributed to the fundamentally different uncertainty quantification mechanisms. A key strength of the CGAN approach lies in its ability to generate physically consistent, high-fidelity synthetic ground motion spectra, making it a promising alternative to standard regression-based GMMs and conventional deep learning models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of conditional generative adversarial network in ground motion modelling encompassing epistemic uncertainty

  • 期刊论文

  • 发表日期：

    2025年10月20日

查看全文
  • Ravi Kanth Sriwastav,
  • Jyothi Yedulla,
  • Pushkal Burde,
  • S. T. G. Raghukanth

摘要

Most existing Ground Motion Models (GMMs) are developed by specifying empirical functional forms derived from expert judgment, with regression techniques used to estimate model coefficients from recorded ground motion data. However, due to the complex and nonlinear nature of earthquake source and path effects, it is difficult to establish a direct analytical relationship between ground motion intensity measures and seismological predictors. This, combined with inherent natural variability, poses significant challenges for accurately characterizing ground motion using predefined functional forms. The emergence of deep learning methods offers a promising avenue for uncovering non-linear correlations among high-dimensional variables. In this study, the deep learning technique, Conditional Generative Adversarial Network (CGAN), is proposed as a novel, data-driven GMM for predicting horizontal-component spectral accelerations over a period range of 0 to 10 s. The model is trained and evaluated using 11675 sets of recorded ground motions from the Engineering Strong motion database (ESM2.0). Model performance is rigorously evaluated through multiple metrics, including residual analysis and comparison with benchmark empirical GMMs. Results demonstrate that the CGAN outperforms traditional models in capturing complex spectral patterns and exhibits superior generalization with reduced prediction bias. Furthermore, a comparative analysis with other deterministic and probabilistic machine learning models developed using the same dataset highlights similarity in aleatory uncertainty but notable differences in epistemic uncertainty estimation, attributed to the fundamentally different uncertainty quantification mechanisms. A key strength of the CGAN approach lies in its ability to generate physically consistent, high-fidelity synthetic ground motion spectra, making it a promising alternative to standard regression-based GMMs and conventional deep learning models.