arXiv subjects

Eric Laloy

Publications and source records attributed to Eric Laloy.

15 recordsLinked to original sources

arXiv2024

Deep learning surrogate for predicting hydraulic conductivity tensors from stochastic discrete fracture-matrix models

Simulating 2D flow in fractured crystalline rock requires 2D stochastic discrete-fracture matrix (DFM) models. To obtain the simulation statistics of interest at an affordable computational cost, we aim to use the multilevel Monte Carlo method. To use this multiscale approach, one needs to upscale the hydraulic conductivity of the fractures by numerical homogenization. In this work, we substitute numerical homogenization with a surrogate model to speed up the computations. In particular, we resort to a deep convolutional neural network (CNN) connected to a deep feed-forward neural network. The equivalent hydraulic conductivity tensor $K_{eq}$ is predicted based on an input spatial random field (SRF) of hydraulic conductivity tensors, cross-section, and hydraulic conductivity of fractures. Three independent surrogates with the same architecture are trained using data from DFM models with three different ratios of hydraulic conductivities of fracture and bulk $K_f/K_b$. As the $K_f/K_b$ ratio increases, the multivariate $K_{eq}$ distribution becomes more complex, and thus, the prediction accuracy of the trained surrogates deteriorates. Regardless of $K_f/K_b$, however, an improvement in the prediction accuracy of the trained surrogates is noted as the considered fracture density of the modeling setup decreases. We also investigate prediction accuracy on input SRFs of different correlation lengths. Upscaling by numerical homogenization and by surrogate modeling is compared on two practical problems: upscaling of the hydraulic conductivity tensor and groundwater flow through a given surface. We obtained equally accurate results for the equivalent hydraulic tensor calculation of upscaled DFM models regardless of the upscaling method. For the groundwater flow problem, the accuracy of quantity of interest imitates the accuracy of $K_{eq}$ predictions.

math.AP↗

arXiv2022

Improving Bayesian radiological profiling of waste drums using Dirichlet priors, Gaussian process priors, and hierarchical modeling

We present three methodological improvements of the "SCK CEN approach" for Bayesian inference of the radionuclide inventory in radioactive waste drums, from radiological measurements. First we resort to the Dirichlet distribution for the prior distribution of the isotopic vector. The Dirichlet distribution possesses the attractive property that the elements of its vector samples sum up to 1. Second, we demonstrate that such Dirichlet priors can be incorporated within an hierarchical modeling of the prior uncertainty in the isotopic vector, when prior information about isotopic composition is available. Our used Bayesian hierarchical modeling framework makes use of this available information but also acknowledges its uncertainty by letting to a controlled extent the information content of the indirect measurement data (i.e., gamma and neutron counts) shape the actual prior distribution of the isotopic vector. Third, we propose to regularize the Bayesian inversion by using Gaussian process (GP) prior modeling when inferring 1D spatially-distributed quantities. As of uncertainty in the efficiencies, we keep using the same stylized drum modeling approach as proposed in our previous work to account for the source distribution uncertainty across the vertical direction of the drum. A series of synthetic tests followed by application to a real waste drum show that combining hierarchical modeling of the prior isotopic composition uncertainty together with GP prior modeling of the vertical Pu profile across the drum works well. We also find that our GP prior can handles both cases with and without spatial correlation. The computational times involved by our proposed approach are on the order of a few hours, say about 2, to provide uncertainty estimates for all variables of interest in the considered inverse problem. This warrants further investigations to speed up the inference.

physics.data-an↗

arXiv2021

Using deep generative neural networks to account for model errors in Markov chain Monte Carlo inversion

Most geophysical inverse problems are nonlinear and rely upon numerical forward solvers involving discretization and simplified representations of the underlying physics. As a result, forward modeling errors are inevitable. In practice, such model errors tend to be either completely ignored, which leads to biased and over-confident inversion results, or only partly taken into account using restrictive Gaussian assumptions. Here, we rely on deep generative neural networks to learn problem-specific low-dimensional probabilistic representations of the discrepancy between high-fidelity and low-fidelity forward solvers. These representations are then used to probabilistically invert for the model error jointly with the target geophysical property field, using the computationally-cheap, low-fidelity forward solver. To this end, we combine a Markov-chain-Monte-Carlo (MCMC) inversion algorithm with a trained convolutional neural network of the spatial generative adversarial network (SGAN) type, whereby at each MCMC step, the simulated low-fidelity forward response is corrected using a proposed model-error realization. Considering the crosshole ground-penetrating radar traveltime tomography inverse problem, we train SGAN networks on traveltime discrepancy images between: (1) curved-ray (high fidelity) and straight-ray (low fidelity) forward solvers; and (2) finite-difference-time-domain (high fidelity) and straight-ray (low fidelity) forward solvers. We demonstrate that the SGAN is able to learn the spatial statistics of the model error and that suitable representations of both the subsurface model and model error can be recovered by MCMC. In comparison with inversion results obtained when model errors are either ignored or approximated by a Gaussian distribution, we find that our method has lower posterior parameter bias and better explains the observed traveltime data.[...]

physics.geo-ph↗

arXiv2021

Speeding up reactive transport simulations in cement systems by surrogate geochemical modeling: deep neural networks and k-nearest neighbors

We accelerate reactive transport (RT) simulation by replacing the geochemical solver in the RT code by a surrogate model or emulator, considering either a trained deep neural network (DNN) or a k-nearest neighbor (kNN) regressor. We focus on 2D leaching of hardened cement paste under diffusive or advective-dispersive transport conditions, a solid solution representation of the calcium silicate hydrates and either 4 or 7 chemical components, and use the HPx (coupled Hydrus-PHREEQC model) reactive transport code as baseline. We find that after training, both our DNN-based and kNN-based codes, HPx-DNN and HPx-kNN, can make satisfactorily to very accurate predictions while providing either a 3 to 9 speedup factor compared to HPx with parallelized geochemical calculations over 4 cores. Benchmarking against single-threaded HPx, these speedup factors become 8 to 33. Overall, HPx-DNN are HPx-kNN are found to achieve a close to optimal speedup when DNN regression and kNN search are performed on a GPU. Importantly, for the more complex 7-components cement system, no emulator that is globally accurate over the full space of possible geochemical conditions could be devised. Instead we therefore build "local" emulators that are only valid over a relevant fraction of the input parameter space. This space is identified by running a coarse and thus computationally cheap full RT simulation, and subsequently explored by kernel density sampling. Future work will focus on improving accuracy for this type of cement systems.

physics.chem-ph↗

arXiv2021

A new framework for experimental design using Bayesian Evidential Learning: the case of wellhead protection area

In this contribution, we predict the wellhead protection area (WHPA, target), the shape and extent of which is influenced by the distribution of hydraulic conductivity (K), from a small number of tracing experiments (predictor). Our first objective is to make stochastic predictions of the WHPA within the Bayesian Evidential Learning (BEL) framework, which aims to find a direct relationship between predictor and target using machine learning. This relationship is learned from a small set of training models (400) sampled from the prior distribution of K. The associated 400 pairs of simulated predictors and targets are obtained through forward modelling. Newly collected field data can then be directly used to predict the approximate posterior distribution of the corresponding WHPA. The uncertainty range of the posterior WHPA distribution is affected by the number and position of data sources (injection wells). Our second objective is to extend BEL to identify the optimal design of data source locations that minimizes the posterior uncertainty of the WHPA. This can be done explicitly, without averaging or approximating because once trained, the BEL model allows the computation of the posterior uncertainty corresponding to any new input data. We use the Modified Hausdorff Distance and the Structural Similarity index metrics to estimate the posterior uncertainty range of the WHPA. Increasing the number of injection wells effectively reduces the derived posterior WHPA uncertainty. Our approach can also estimate which injection wells are more informative than others, as validated through a k-fold cross-validation procedure. Overall, the application of BEL to experimental design makes it possible to identify the data sources maximizing the information content of any measurement data.

cs.LG↗

arXiv2021

Bayesian inference of 1D activity profiles from segmented gamma scanning of a heterogeneous radioactive waste drum

We present a Bayesian approach to probabilistically infer vertical activity profiles within a radioactive waste drum from segmented gamma scanning (SGS) measurements. Our approach resorts to Markov chain Monte Carlo (MCMC) sampling using the state-of-the-art Hamiltonian Monte Carlo (HMC) technique and accounts for two important sources of uncertainty: the measurement uncertainty and the uncertainty in the source distribution within the drum. In addition, our efficiency model simulates the contributions of all considered segments to each count measurement. Our approach is first demonstrated with a synthetic example, after which it is used to resolve the vertical activity distribution of 5 nuclides in a real waste package.

physics.data-an↗

arXiv2020

Approaching geoscientific inverse problems with vector-to-image domain transfer networks

We present vec2pix, a deep neural network designed to predict categorical or continuous 2D subsurface property fields from one-dimensional measurement data (e.g., time series), thereby offering a new approach to solve inverse problems. The performance of the method is investigated through two types of synthetic inverse problems: (a) a crosshole ground penetrating radar (GPR) tomography experiment with GPR travel times being used to infer a 2D velocity field, and (2) a multi-well pumping experiment within an unconfined aquifer with time series of transient hydraulic heads being used to retrieve a 2D hydraulic conductivity field. For each type of problem, both a multi-Gaussian and a binary channelized subsurface domain with long-range connectivity are considered. Using a training set of 20,000 examples (implying as many forward model evaluations), the method is found to recover a 2D model that is in much closer agreement with the true model than the closest training model in the forward-simulated data space. Further testing with smaller training sample sizes shows only a moderate reduction in performance when using 5000 training examples only. Even if the recovered models are visually close to the true ones, the data misfits associated with their forward responses are generally larger than the noise level used to contaminate the true data. Uncertainty of the inverse solution is partially assessed using deep ensembles, in which the network is trained repeatedly with random initialization. Overall, this study advances understanding of how to use deep learning to infer subsurface models from indirect measurement data.

physics.geo-ph↗

arXiv2020

Deep generative models in inversion: a review and development of a new approach based on a variational autoencoder

When solving inverse problems in geophysical imaging, deep generative models (DGMs) may be used to enforce the solution to display highly structured spatial patterns which are supported by independent information (e.g. the geological setting) of the subsurface. In such case, inversion may be formulated in a latent space where a low-dimensional parameterization of the patterns is defined and where Markov chain Monte Carlo or gradient-based methods may be applied. However, the generative mapping between the latent and the original (pixel) representations is usually highly nonlinear which may cause some difficulties for inversion, especially for gradient-based methods. In this contribution we review the conceptual framework of inversion with DGMs and study the principal causes of the nonlinearity of the generative mapping. As a result, we identify a conflict between two goals: the accuracy of the generated patterns and the feasibility of gradient-based inversion. In addition, we show how some of the training parameters of a variational autoencoder, which is a particular instance of a DGM, may be chosen so that a tradeoff between these two goals is achieved and acceptable inversion results are obtained with a stochastic gradient-descent scheme. A test case using truth models with channel patterns of different complexity and cross-borehole traveltime tomographic data involving both a linear and a nonlinear forward operator is used to assess the performance of the proposed approach.

physics.geo-ph↗

arXiv2020

Pixel-wise Conditioned Generative Adversarial Networks for Image Synthesis and Completion

Generative Adversarial Networks (GANs) have proven successful for unsupervised image generation. Several works have extended GANs to image inpainting by conditioning the generation with parts of the image to be reconstructed. Despite their success, these methods have limitations in settings where only a small subset of the image pixels is known beforehand. In this paper we investigate the effectiveness of conditioning GANs when very few pixel values are provided. We propose a modelling framework which results in adding an explicit cost term to the GAN objective function to enforce pixel-wise conditioning. We investigate the influence of this regularization term on the quality of the generated images and the fulfillment of the given pixel constraints. Using the recent PacGAN technique, we ensure that we keep diversity in the generated samples. Conducted experiments on FashionMNIST show that the regularization term effectively controls the trade-off between quality of the generated images and the conditioning. Experimental evaluation on the CIFAR-10 and CelebA datasets evidences that our method achieves accurate results both visually and quantitatively in term of Fréchet Inception Distance, while still enforcing the pixel conditioning. We also evaluate our method on a texture image generation task using fully-convolutional networks. As a final contribution, we apply the method to a classical geological simulation application.

eess.IV↗

arXiv2019

Pixel-wise Conditioning of Generative Adversarial Networks

Generative Adversarial Networks (GANs) have proven successful for unsupervised image generation. Several works extended GANs to image inpainting by conditioning the generation with parts of the image one wants to reconstruct. However, these methods have limitations in settings where only a small subset of the image pixels is known beforehand. In this paper, we study the effectiveness of conditioning GANs by adding an explicit regularization term to enforce pixel-wise conditions when very few pixel values are provided. In addition, we also investigate the influence of this regularization term on the quality of the generated images and the satisfaction of the conditions. Conducted experiments on MNIST and FashionMNIST show evidence that this regularization term allows for controlling the trade-off between quality of the generated images and constraint satisfaction.

cs.CV↗

arXiv2019

Emulation of CPU-demanding reactive transport models: comparison of Gaussian processes, polynomial chaos expansion and deep neural networks

This paper presents a detailed comparison between 3 methods for emulating CPU-intensive reactive transport models (RTMs): Gaussian processes (GPs), polynomial chaos expansion (PCE) and deep neural networks (DNNs). Besides direct emulation of the simulated uranium concentration time series, replacing the original RTM by its emulator is also investigated for global sensitivity analysis (GSA), uncertainty propagation (UP) and probabilistic calibration using Markov chain Monte Carlo (MCMC) sampling. The selected DNN is found to be superior to both GPs and PCE in reproducing the input - output behavior of the considered 8-dimensional and 13-dimensional CPU-intensive RTMs. Furthermore, the two used PCE variants: standard PCE and sparse PCE (sPCE) appear to always provide the least accuracy while not differing much in performance. As a consequence of its better emulation capabilities, the used DNN outperforms the two other methods for UP. In addition, DNNs and GPs offer equally good approximations to the true first-order and total-order Sobol sensitivity indices while PCE does somewhat less well. Most surprisingly, despite its superior emulation skills the DNN approach leads to the worst solution of the considered synthetic inverse problem which involves 1224 measurement data with low noise. This apparently contradicting behavior is at least partially due to the small but complicated deterministic noise that affects the DNN-based predictions. Indeed, this complex error structure can drive the emulated solutions far away from the true posterior distribution. Overall, our findings indicate that when the available training set is relatively small (75 - to 500 input - output examples) and fixed beforehand, DNNs can well emulate RTMs but are not suited to emulation-based inversion. In contrast, GPs perform fairly well across all considered tasks: direct emulation, GSA, UP, and inversion.

physics.comp-ph↗

arXiv2019

Gradient-based deterministic inversion of geophysical data with Generative Adversarial Networks: is it feasible?

Global probabilistic inversion within the latent space learned by a Generative Adversarial Network (GAN) has been recently demonstrated. Compared to inversion on the original model space, using the latent space of a trained GAN can offer the following benefits: (1) the generated model proposals are geostatistically consistent with the prescribed prior training image (TI), and (2) the parameter space is reduced by orders of magnitude compared to the original model space. Nevertheless, exploring the learned latent space by state-of-the-art Markov chain Monte Carlo (MCMC) methods may still require a large computational effort. As an alternative, parameters in this latent space could possibly be optimized with much less computationally expensive gradient-based methods. We show that due to the typically highly nonlinear relationship between the latent space and the associated output space of a GAN, gradient-based deterministic inversion may fail even when considering a linear forward physical model. We tested two deterministic inversion approaches: a quasi-Newton gradient descent using the Adam algorithm and a Gauss-Newton (GN) method that makes use of the Jacobian matrix calculated by finite-differencing. For a channelized binary TI and a synthetic linear crosshole ground penetrating radar (GPR) tomography problem involving 576 measurements with low noise, we observe that when allowing for a total of 10,000 iterations only 13% of the gradient descent trials locate a solution that has the required data misfit. The tested GN inversion was unable to recover a solution with the appropriate data misfit. Our results suggest that deterministic inversion performance strongly depends on the inversion approach, starting model, true reference model, number of iterations and noise realization. In contrast, computationally-expensive probabilistic global optimization always finds an appropriate solution.

physics.geo-ph↗

arXiv2019

Dilated Spatial Generative Adversarial Networks for Ergodic Image Generation

Generative models have recently received renewed attention as a result of adversarial learning. Generative adversarial networks consist of samples generation model and a discrimination model able to distinguish between genuine and synthetic samples. In combination with convolutional (for the discriminator) and de-convolutional (for the generator) layers, they are particularly suitable for image generation, especially of natural scenes. However, the presence of fully connected layers adds global dependencies in the generated images. This may lead to high and global variations in the generated sample for small local variations in the input noise. In this work we propose to use architec-tures based on fully convolutional networks (including among others dilated layers), architectures specifically designed to generate globally ergodic images, that is images without global dependencies. Conducted experiments reveal that these architectures are well suited for generating natural textures such as geologic structures .

cs.CV↗

arXiv2019

Training-image based geostatistical inversion using a spatial generative adversarial neural network

Probabilistic inversion within a multiple-point statistics framework is often computationally prohibitive for high-dimensional problems. To partly address this, we introduce and evaluate a new training-image based inversion approach for complex geologic media. Our approach relies on a deep neural network of the generative adversarial network (GAN) type. After training using a training image (TI), our proposed spatial GAN (SGAN) can quickly generate 2D and 3D unconditional realizations. A key characteristic of our SGAN is that it defines a (very) low-dimensional parameterization, thereby allowing for efficient probabilistic inversion using state-of-the-art Markov chain Monte Carlo (MCMC) methods. In addition, available direct conditioning data can be incorporated within the inversion. Several 2D and 3D categorical TIs are first used to analyze the performance of our SGAN for unconditional geostatistical simulation. Training our deep network can take several hours. After training, realizations containing a few millions of pixels/voxels can be produced in a matter of seconds. This makes it especially useful for simulating many thousands of realizations (e.g., for MCMC inversion) as the relative cost of the training per realization diminishes with the considered number of realizations. Synthetic inversion case studies involving 2D steady-state flow and 3D transient hydraulic tomography with and without direct conditioning data are used to illustrate the effectiveness of our proposed SGAN-based inversion. For the 2D case, the inversion rapidly explores the posterior model distribution. For the 3D case, the inversion recovers model realizations that fit the data close to the target level and visually resemble the true model well.

stat.ML↗

arXiv2017

Inversion using a new low-dimensional representation of complex binary geological media based on a deep neural network

Efficient and high-fidelity prior sampling and inversion for complex geological media is still a largely unsolved challenge. Here, we use a deep neural network of the variational autoencoder type to construct a parametric low-dimensional base model parameterization of complex binary geological media. For inversion purposes, it has the attractive feature that random draws from an uncorrelated standard normal distribution yield model realizations with spatial characteristics that are in agreement with the training set. In comparison with the most commonly used parametric representations in probabilistic inversion, we find that our dimensionality reduction (DR) approach outperforms principle component analysis (PCA), optimization-PCA (OPCA) and discrete cosine transform (DCT) DR techniques for unconditional geostatistical simulation of a channelized prior model. For the considered examples, important compression ratios (200 - 500) are achieved. Given that the construction of our parameterization requires a training set of several tens of thousands of prior model realizations, our DR approach is more suited for probabilistic (or deterministic) inversion than for unconditional (or point-conditioned) geostatistical simulation. Probabilistic inversions of 2D steady-state and 3D transient hydraulic tomography data are used to demonstrate the DR-based inversion. For the 2D case study, the performance is superior compared to current state-of-the-art multiple-point statistics inversion by sequential geostatistical resampling (SGR). Inversion results for the 3D application are also encouraging.

stat.ML↗