Tsubasa Onishi, Hongquan Chen, A. Datta-Gupta, Srikanta Mishra
{"title":"结合简化物理模型的高效深度学习工作流程用于非常规油藏地下成像","authors":"Tsubasa Onishi, Hongquan Chen, A. Datta-Gupta, Srikanta Mishra","doi":"10.2118/206065-ms","DOIUrl":null,"url":null,"abstract":"\n We present a novel deep learning-based workflow incorporating a reduced physics model that can efficiently visualize well drainage volume and pressure front propagation in unconventional reservoirs in near real-time. The visualizations can be readily used for qualitative and quantitative characterization and forecasting of unconventional reservoirs.\n Our aim is to develop an efficient workflow that allows us to ‘see’ within the subsurface given measured data, such as production data. The most simplistic way to achieve the goal will be to merely train a deep learning-based regression model where the input consists of some measured data, and the output is a subsurface image, such as pressure field. However, the high output dimension that corresponds to spatio-temporal steps makes the training inefficient. To address this challenge, an autoencoder network is applied to discover lower dimensional latent variables that represent high dimensional output images. In our approach, the regression model is trained to predict latent variables, instead of directly constructing an image. In the prediction step, the trained regression model first predicts latent variables given measured data, then the latent variables will be used as inputs of the trained decoder to generate a subsurface image. In addition, fast marching-method (FMM)-based rapid simulation workflow which transforms original 2D or 3D problems into 1D problems is used in place of full-physics simulation to efficiently generate datasets for training. The capability of the FMM-based rapid simulation allows us to generate sufficient datasets within realistic simulation times, even for field scale applications.\n We first demonstrate the proposed approach using a simple illustrative example. Next, the approach is applied to a field scale reservoir model built after the publicly available data on the Hydraulic Fracturing Test Site-I (HFTS-I), which is sufficiently complex to demonstrate the power and efficacy of the approach. We will further demonstrate the utility of the approach to account for subsurface uncertainty. Our approach, for the first time, allows data-driven visualization of unconventional well drainage volume in 3D. The novelty of our approach is the framework which combines the strengths of deep learning-based models and the FMM-based rapid simulation. The workflow has flexibility to incorporate various spatial and temporal data types.","PeriodicalId":10965,"journal":{"name":"Day 3 Thu, September 23, 2021","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2021-09-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"1","resultStr":"{\"title\":\"An Efficient Deep Learning-Based Workflow Incorporating a Reduced Physics Model for Subsurface Imaging in Unconventional Reservoirs\",\"authors\":\"Tsubasa Onishi, Hongquan Chen, A. Datta-Gupta, Srikanta Mishra\",\"doi\":\"10.2118/206065-ms\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"\\n We present a novel deep learning-based workflow incorporating a reduced physics model that can efficiently visualize well drainage volume and pressure front propagation in unconventional reservoirs in near real-time. The visualizations can be readily used for qualitative and quantitative characterization and forecasting of unconventional reservoirs.\\n Our aim is to develop an efficient workflow that allows us to ‘see’ within the subsurface given measured data, such as production data. The most simplistic way to achieve the goal will be to merely train a deep learning-based regression model where the input consists of some measured data, and the output is a subsurface image, such as pressure field. However, the high output dimension that corresponds to spatio-temporal steps makes the training inefficient. To address this challenge, an autoencoder network is applied to discover lower dimensional latent variables that represent high dimensional output images. In our approach, the regression model is trained to predict latent variables, instead of directly constructing an image. In the prediction step, the trained regression model first predicts latent variables given measured data, then the latent variables will be used as inputs of the trained decoder to generate a subsurface image. In addition, fast marching-method (FMM)-based rapid simulation workflow which transforms original 2D or 3D problems into 1D problems is used in place of full-physics simulation to efficiently generate datasets for training. The capability of the FMM-based rapid simulation allows us to generate sufficient datasets within realistic simulation times, even for field scale applications.\\n We first demonstrate the proposed approach using a simple illustrative example. Next, the approach is applied to a field scale reservoir model built after the publicly available data on the Hydraulic Fracturing Test Site-I (HFTS-I), which is sufficiently complex to demonstrate the power and efficacy of the approach. We will further demonstrate the utility of the approach to account for subsurface uncertainty. Our approach, for the first time, allows data-driven visualization of unconventional well drainage volume in 3D. The novelty of our approach is the framework which combines the strengths of deep learning-based models and the FMM-based rapid simulation. The workflow has flexibility to incorporate various spatial and temporal data types.\",\"PeriodicalId\":10965,\"journal\":{\"name\":\"Day 3 Thu, September 23, 2021\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2021-09-15\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"1\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Day 3 Thu, September 23, 2021\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.2118/206065-ms\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Day 3 Thu, September 23, 2021","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.2118/206065-ms","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
An Efficient Deep Learning-Based Workflow Incorporating a Reduced Physics Model for Subsurface Imaging in Unconventional Reservoirs
We present a novel deep learning-based workflow incorporating a reduced physics model that can efficiently visualize well drainage volume and pressure front propagation in unconventional reservoirs in near real-time. The visualizations can be readily used for qualitative and quantitative characterization and forecasting of unconventional reservoirs.
Our aim is to develop an efficient workflow that allows us to ‘see’ within the subsurface given measured data, such as production data. The most simplistic way to achieve the goal will be to merely train a deep learning-based regression model where the input consists of some measured data, and the output is a subsurface image, such as pressure field. However, the high output dimension that corresponds to spatio-temporal steps makes the training inefficient. To address this challenge, an autoencoder network is applied to discover lower dimensional latent variables that represent high dimensional output images. In our approach, the regression model is trained to predict latent variables, instead of directly constructing an image. In the prediction step, the trained regression model first predicts latent variables given measured data, then the latent variables will be used as inputs of the trained decoder to generate a subsurface image. In addition, fast marching-method (FMM)-based rapid simulation workflow which transforms original 2D or 3D problems into 1D problems is used in place of full-physics simulation to efficiently generate datasets for training. The capability of the FMM-based rapid simulation allows us to generate sufficient datasets within realistic simulation times, even for field scale applications.
We first demonstrate the proposed approach using a simple illustrative example. Next, the approach is applied to a field scale reservoir model built after the publicly available data on the Hydraulic Fracturing Test Site-I (HFTS-I), which is sufficiently complex to demonstrate the power and efficacy of the approach. We will further demonstrate the utility of the approach to account for subsurface uncertainty. Our approach, for the first time, allows data-driven visualization of unconventional well drainage volume in 3D. The novelty of our approach is the framework which combines the strengths of deep learning-based models and the FMM-based rapid simulation. The workflow has flexibility to incorporate various spatial and temporal data types.