基于物理过程的CO2+O2地浸采铀反应性溶质运移模型需要采用高分辨率网格来捕获介质物理-化学参数空间变异性和对流-弥散-化学反应多过程耦合特征，往往面临巨大的模型计算挑战。传统的替代模型处理高维数据空间分布预测时面临精度和维数诅咒问题，基于条件深度卷积生成对抗网络(conditional Deep Convolutional Generative Adversarial Network，cDC-GAN)，本文提出了一种多重输入图像到输出图像回归的 cDC-GAN 替代建模组件，建立了高维物理-化学非均质场(渗透率场和铀矿物品位场)和铀浸出浓度分布场之间的映射关系，训练集和测试集样品中结构相似性指数的中值达0。98以上，可以作为砂岩型铀矿CO2+O2溶浸过程数值模型的替代方案。cDC-GAN替代建模不受底层物理模型的限制，进而可为复杂反应性溶质运移模型的参数识别、不确定性分析、全局敏感性分析和模拟优化方案设计等提供通用框架。
Predicting CO2+O2 in-situ leaching process in physico-chemical heterogeneous sandstone-type uranium ore using a cDC-GAN-based proxy model
A CO2+O2 in-situ leaching uranium reactive solute transport model based on physical processes requires high-resolution grids to capture the spatial variability of medium physical-chemical parameters and the multi-process coupling characteristics of advection-dispersion-chemical reactions.This often faces significant computational challenges.Traditional surrogate models encounter precision and curse of dimensionality issues when predicting high-dimensional data spatial distributions.In this paper,a conditional Deep Convolutional Generative Adversarial Network(cDC-GAN)is proposed as a surrogate modeling component for multi-input image to output image regression.The mapping relationship between high-dimensional physical-chemical heterogeneous fields(permeability fields and uranium mineral grade fields)and uranium leaching concentration distribution fields is established.The median structural similarity index in the training and test sample sets exceeds 0.98,making it a viable alternative to numerical models of the CO2+O2 in-situ leaching process for sandstone-hosted uranium deposits.The cDC-GAN surrogate modeling is not constrained by the underlying physical model,thereby providing a general framework for parameter identification,uncertainty analysis,global sensitivity analysis,and simulation optimization design of complex reactive solute transport models.
CO2+O2 in-situ leaching of uraniumdeep learningimage predictionuncertainty analysis
万方数据
