Scholarlink

<p>In history matching workflows, facies parameters cannot be straightforwardly calibrated by ensemble-based methods, because they are characterized by a categorical nature. In this work, we treat 3D synthetic reservoir images with channelized facies adopting a Convolutional Variational AutoEncoder (CVAE), which is then coupled with an Ensemble Smoother with Multiple Data Assimilation (ESMDA) workflow. The goal is to parameterize the ensemble of reservoir models into a latent space and to calibrate it. An extensive hyperparamters tuning is conducted: the identification of optimal parameters allows to find a compromise between the reconstruction accuracy and the generative ability of the network. Thanks to this study, the results of the history match are significantly improved. A sensitivity analysis to assess the possibility of reducing the training set size without deteriorating the ESMDA accuracy is also carried out.</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

Investigation of convolutional variational autoencoders for facies calibration within ensemble-smoother multiple data assimilation

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Emanuela Abbate,
  • Laura Dovera,
  • Davide Di Curzio,
  • Andrea Da Pra

摘要

In history matching workflows, facies parameters cannot be straightforwardly calibrated by ensemble-based methods, because they are characterized by a categorical nature. In this work, we treat 3D synthetic reservoir images with channelized facies adopting a Convolutional Variational AutoEncoder (CVAE), which is then coupled with an Ensemble Smoother with Multiple Data Assimilation (ESMDA) workflow. The goal is to parameterize the ensemble of reservoir models into a latent space and to calibrate it. An extensive hyperparamters tuning is conducted: the identification of optimal parameters allows to find a compromise between the reconstruction accuracy and the generative ability of the network. Thanks to this study, the results of the history match are significantly improved. A sensitivity analysis to assess the possibility of reducing the training set size without deteriorating the ESMDA accuracy is also carried out.