Statistical Machine Learning Based Geological Inverse Modeling And Its Application

Geostatistics modeling is one of the major approaches to produce random simulated realizations in earth science.Among them,multiple-point geostatistics plays an important role in characterizing complex subsurface structure.But it suffers from a set of problems(e.g.,how to select an applicable training image,high computational cost).The uncertainty of Training Image can cause a great impact to the following studies based on multiple-point geostatistics.However,only a few studies have paid attention to how to choose a proper training image for multiple-point geostatistics.On the other hand,geological subsurface inversion plays an important role in building subsurface models,since there are more state parameter data(e.g.hydraulic head)than the physical hard data(e.g.hydraulic conductivity).However,geological inverse process is often based on the prior conceptual model(e.g.training image in multiple-point geostatistics).In real scenarios,as the geological subsurface structure is complex and diverse,it is difficult to select an applicable prior conceptual model,which contributes a large uncertainty in the inversion process.Therefore,it is necessary to take the uncertainty of conceptual model into account.To address the above problems,we propose machine learning based model selection and multivariate parameter estimation approaches:(1)A Recurrent Neural Networks(RNNs)based model selection method to select an applicable training image for multiple-point geostatistics according to state parameter data(hydraulic head time series).(2)A Conditional Wasserstein Generative Adversarial Network(CWGAN)based method to estimate physical hard data conditioned on state parameter data under multiple conceptual models.The results show that:(1)The proposed RNNs based method can have a 97.63% accuracy in training image selection task.(2)The physical connection between state parameters and physical data can be learned by CWGAN.And the trained model can inversely reconstruct the geological subsurface structure using observations of hydraulic head time series.(3)CWGAN has a promising performance on the reconstruction of hydraulic conductivity fields with a Mean Accuracy over 80%and a Structural Similarity(SSIM)over 0.6 on both training set and testing set.