Maintenance of aging buried infrastructure and reinforced concrete are critical issues in the United States. Inexpensive non-destructive techniques for mapping and imaging infrastructure and defects are an integral component of maintenance. Ground penetrating radar (GPR) is a widely-used non-destructive tool for locating buried infrastructure and for imaging rebar and other features of interest to civil engineers. Conventional acquisition and interpretation of GPR profiles is based on the arrival times of strong reflected/diffracted returns, and qualitative interpretation of return amplitudes. Features are thereby generally well located, but their material properties are only qualitatively assessed. For example, in the typical imaging of buried pipes, the average radar wave velocity through the overlying soil is estimated, but the properties of the pipe itself are not quantitatively resolved. For pipes on the order of the radar wavelength (<5-35 cm), pipe dimensions and infilling material remain ambiguous. Full waveform inversion (FWI) methods exploit the entire radar return rather than the time and peak amplitude. FWI can generate better quantitative estimates of subsurface properties. In recent decades FWI methods, developed for seismic oil exploration, have been adapted and advanced for GPR with encouraging results. To date, however, FWI methods for GPR data have not been specifically tuned and applied on surface collected common offset GPR data, which are the most common type of GPR data for engineering applications. I present an effective FWI method specifically tailored for common-offset GPR data. This method is composed of three main components, the forward modeling, wavelet estimation and inversion tools. For the forward modeling and iterative data inversion I use two open-source software packages, gprMax and PEST. The source wavelet, which is the most challenging component that guarantees the success of the method, is estimated with a novel Sparse Blind Deconvolution (SBD) algorithm that I have developed. The present dissertation indicates that with FWI, GPR can yield better quantitative estimates, for example, of both the diameters of small pipes and rebar and their electromagnetic properties (permittivity, conductivity). Also better estimates of electrical properties of the surrounding media (i.e. soil or concrete) are achieved with FWI.

京都大学 / 0048 / 新制・課程博士 / 博士(工学) / 甲第20061号 / 工博第4249号 / 新制||工||1658(附属図書館) / 京都大学大学院工学研究科社会基盤工学専攻 / (主査)教授 三ｹ田 均, 教授 小池 克明, 教授 木村 亮 高梨 将 / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DFAM

The main objective of this thesis is to develop novel misfit functions for full-waveform inversion such that (a) the estimation of the long-wavelength model will less likely stagnate in spurious local minima and (b) the inversion is immune to wavelet inaccuracy. First, I investigate the pros and cons of misfit functions based on optimal transport theory to indicate the traveltime discrepancy for seismic data. Even though the mathematically well-defined optimal transport theory is robust to highlight the traveltime difference between two probability distributions, it becomes restricted as applied to seismic data mainly because the seismic data are not probability distribution functions. We then develop a misfit function combining the local cross-correlation and dynamic time warping. This combination enables the proposed misfit automatically identify arrivals associated with a phase shift. Numerical and field data examples demonstrate its robustness for early arrivals and limitations for later arrivals.%, which means that a proper pre-processing step is still required. Next, we introduce differentiable dynamic time warping distance as the misfit function highlighting the traveltime discrepancy without non-trivial human intervention. Compared to the conventional warping distance, the differentiable version retains the property of representing the traveltime difference; moreover, it can eliminate abrupt changes in the adjoint source, which helps full-waveform inversion converge to geologically relevant estimates. Finally, we develop a misfit function entitled the deconvolutional double-difference measurement. The new misfit measures the first difference by deconvolution rather than cross-correlation. We also present the derivation of the adjoint source with the new misfit function. Numerical examples and mathematical proof demonstrate that this modification makes full-waveform inversion with the deconvolutional double-difference measurement immune to wavelet inaccuracy.

Full Waveform Inversion (FWI) is a widely used optimization technique for subsurface imaging where the goal is to estimate the seismic wave velocity beneath the Earth's surface from the observed seismic data at the surface. The problem is primarily governed by the wave equation, which is a non-linear second-order partial differential equation. A number of approaches have been developed for FWI including physics-based iterative numerical solvers as well as data-driven machine learning (ML) methods. Existing numerical solutions to FWI suffer from three major challenges: (1) sensitivity to initial velocity guess (2) non-convex loss landscape, and (3) sensitivity to noise. Additionally, they suffer from high computational cost, making them infeasible to apply in complex real-world applications. Existing ML solutions for FWI only solve for the inverse and are prone to yield non-unique solutions. In this work, we propose to solve both forward and inverse problems jointly to alleviate the issue of non-unique solutions for an inverse problem. We study the FWI problem from a new perspective and propose a novel approach based on Invertible Neural Networks. This type of neural network is designed to learn bijective mappings between the input and target distributions and hence they present a potential solution to solve forward and inverse problems jointly. In this thesis, we developed a data-driven framework that can be used to learn forward and inverse mappings between any arbitrary input and output space. Our model, Invertible X-net, can be used to solve FWI to obtain high-quality velocity images and also predict the seismic waveforms data. We compare our model with the existing baseline mod- els and show that our model outperforms them in velocity reconstruction on the OpenFWI dataset. Additionally, we also compare the predicted waveforms with a baseline and ground truth and show that our model is capable of predicting highly accurate seismic waveforms simultaneously. / Master of Science / Recent advancements in deep learning have led to the development of sophisticated methods that can be used to solve scientific problems in many disciplines including medical imaging, geophysics, and signal processing. For example, in geophysics, we study the internal structure of the Earth from indirect physical measurements. Often, these kind of problems are challenging due to existence of non-unique and unstable solutions. In this thesis, we look at one such problem called Full Waveform Inversion which aims to estimate velocity of mechanical wave inside the Earth from wave amplitude observations on the surface. For this problem, we explore a special class of neural networks that allows to uniquely map the input and output space and thus alleviate the non-uniqueness and instability in performing Full Waveform Inversion for seismic imaging.

Full Waveform Inversion (FWI) is a powerful seismic imaging technique used to reconstruct high-resolution velocity models of the subsurface. It relies on the inversion of seismic data acquired from multiple sources and receivers to estimate the mechanical properties of geologic materials and can be used to detect anomalous subsurface conditions. The accuracy of FWI results is influenced by various factors related to the workflow used for its implementation. This includes the survey parameters, the mathematical framework of the inversion, and the complexity of the subsurface conditions modeled during the inversion process. Therefore, it is crucial to have a fundamental understanding of the interplay between these factors and their impact on the accuracy of the reconstructed model, particularly given the effects of these factors on computational costs. This is an area that has been understudied within the context of near-surface geotechnical applications for anomaly detection, which is an application that presents unique challenges relative to seismic exploration for hydrocarbons where FWI has been more fully developed. One key aspect that has not received sufficient attention is the impact of survey parameters on the accuracy of FWI results. The lack of formal research in this topic may lead to near-surface FWI studies that use more seismic sources than required for subsurface feature reconstruction, which results in data collection and computational inefficiencies. The selection of misfit function and starting model are also essential factors influencing the reliability of the reconstructed model. The physics employed for forward modeling can also affect the ability to simulate wave propagation in the domain of interest. These factors have significant implications for near-surface applications of FWI, and further research is required to explore their interplay and improve FWI workflow.Given the gaps in the current implementation of FWI for geotechnical applications, this research will explore the role of parameterization and workflow on FWI results when applied to anomaly detection in karst conditions. This will include selection of an FWI workflow that can improve the feasibility of fieldwork and reduce the processing time. The research will investigate four key factors of the FWI workflow (i.e., survey design, initial model, misfit function, and forward modeling physics) for detection of sinkholes using numerical and field testing in different subsurface conditions. Overall, the outcomes of this research will help practitioners with more appropriate choices in the FWI process and consequently promote its high potential in near-surface applications. / Civil Engineering

Drilled shafts are deep foundation elements created by excavating cylindrical shafts into the ground and filling them with concrete. Given the types of structures they support, failure to meet their performance criteria can jeopardize public safety and cause severe financial losses. Consequently, quality control measures are warranted to ensure these foundations meet design specifications, particularly with respect to their structural integrity and geotechnical capacity. Due to their inaccessibility, non-destructive testing (NDT) techniques have received much attention for drilled shaft quality control. However, there are limitations in the NDT tools currently used for structural integrity testing. Moreover, there is no current NDT tool to evaluate conditions underlying drilled shaft excavations and aid in verifying geotechnical capacity. The main objective of this research is to examine the development of new NDT methodologies to address some of the limitations in the inspection of drilled shaft structural integrity and geotechnical conditions underlying their excavations. The use of stress waves in large laboratory models is first examined to evaluate the performance of ray-based techniques for detecting anomalies. The study then continues to investigate the improvements offered by using a full waveform inversion (FWI) approach to analyze the stress wave data. A hybrid, multi-scale FWI workflow is recommended to increase the chance of the convergence of the inversion algorithms. Additionally, the benefits of a multi-parameter FWI are discussed. Since FWI is computationally expensive, a sequential optimal experimental design (SOED) analysis is proposed to determine the optimal hardware configurations for each application. The resulting benefit-cost curves from this analysis allow for designing an NDT survey that matches the available resources for the project. / Civil Engineering

We are concerned with the high-fidelity subsurface imaging of the soil, which commonly arises in geotechnical site characterization and geophysical explorations. Specifically, we attempt to image the spatial distribution of the Lame parameters in semi-infinite, three-dimensional, arbitrarily heterogeneous formations, using surficial measurements of the soil's response to probing elastic waves. We use the complete waveforms of the medium's response to drive the inverse problem. Specifically, we use a partial-differential-equation (PDE)-constrained optimization approach, directly in the time-domain, to minimize the misfit between the observed response of the medium at select measurement locations, and a computed response corresponding to a trial distribution of the Lame parameters. We discuss strategies that lend algorithmic robustness to the proposed inversion schemes. To limit the computational domain to the size of interest, we employ perfectly-matched-layers (PMLs). The PML is a buffer zone that surrounds the domain of interest, and enforces the decay of outgoing waves. In order to resolve the forward problem, we present a hybrid finite element approach, where a displacement-stress formulation for the PML is coupled to a standard displacement-only formulation for the interior domain, thus leading to a computationally cost-efficient scheme. We discuss several time-integration schemes, including an explicit Runge-Kutta scheme, which is well-suited for large-scale problems on parallel computers. We report numerical results demonstrating stability and efficacy of the forward wave solver, and also provide examples attesting to the successful reconstruction of the two Lame parameters for both smooth and sharp profiles, using synthetic records. We also report the details of two field experiments, whose records we subsequently used to drive the developed inversion algorithms in order to characterize the sites where the field experiments took place. We contrast the full-waveform-based inverted site profile against a profile obtained using the Spectral-Analysis-of-Surface-Waves (SASW) method, in an attempt to compare our methodology against a widely used concurrent inversion approach. We also compare the inverted profiles, at select locations, with the results of independently performed, invasive, Cone Penetrometer Tests (CPTs). Overall, whether exercised by synthetic or by physical data, the full-waveform inversion method we discuss herein appears quite promising for the robust subsurface imaging of near-surface deposits in support of geotechnical site characterization investigations.

L'inversion des formes d'onde (full waveform inversion, FWI) a suscité un intérêt dans le monde entier pour sa capacité à estimer de manière précise et détaillée les propriétés physiques du sous-sol. La FWI est généralement formulée sous la forme d'un problème d'ajustement des données par moindres carrés et résolus par une approche linéarisée utilisant des méthodes d'optimisation locales. Cependant, la FWI est bien connue de souffrir du problème de saut de phase rendant les résultats fortement dépendant de la qualité des modèles initiaux. L'inversion des formes d'ondes des arrivées réfléchies (reflection waveform inversion, RWI) a récemment été proposée pour atténuer ce problème en supposant une séparation d'échelle entre le modèle de vitesse lisse et le modèle de réflectivité à haut nombre d'onde. La formulation de RWI considère explicitement les ondes réfléchies afin d'extraire de ces ondes une information sur les variations lisses de vitesse des zones profondes. Cependant, la méthode néglige les ondes transmises qui contraignant les informations lisses de vitesse en proche surface.Dans cette thèse, une étude de la sensibilité en nombre d'ondes des méthodes de FWI et RWI a d'abord été revisitée dans le cadre de la tomographie en diffraction et des décompositions orthogonales. A partir de cette analyse, je propose une nouvelle méthode, à savoir l'inversion jointe des formes d'ondes transmises et réfléchies (joint full waveform inversion, JFWI). La méthode propose une formulation unifiée pour combiner la FWI des transmissions et la RWI pour les réflexions, donnant naturellement une sensibilité commune aux petits nombres d'onde venant des arrivées grand-angle et réfléchies. Les composantes à hauts nombres d'onde sont naturellement atténuées par la formulation. Pour satisfaire l'hypothèse de séparation d'échelle, j'utilise une paramétrisation du sous-sol basée sur la vitesse des ondes de compression et l'impédance acoustique. La complexité temporelle de cette approche est le double de la méthode de FWI classique et la requête mémoire reste la même.Une procédure d'inversion est ensuite proposée, permettant d'estimer alternativement le modèle de la vitesse du sous-sol par JFWI et l'impédance inversion de formes d'ondes réfléchies. Un exemple synthétique réaliste du modèle de Valhall est d'abord utilisé avec des données de streamer et à partir d'un modèle initial très lisse. Dans ce cadre, alors que la FWI converge vers un minimum local, la JFWI réussit à reconstruire un modèle de vitesse lisse de bonne qualité. La prise en compte des ondes tournante par la JFWI montre un fort intérêt pour la qualité de reconstruction superficielle, comparée à la méthode RWI seule. Cela se traduit ensuite par une reconstruction améliorée en profondeur. Le modèle de vitesse lisse construit par JFWI peut ensuite être considéré comme modèle initial pour la FWI classique, afin d'injecter le contenu en haut nombres d'onde tout en évitant le problème de saut de phase.Les avantages et limites de l'approche de JFWI sont ensuite étudiés dans une application sur données réelles, venant d'un profil 2D de données de fond de mer (OBC) recoupant un nuage de gaz au dessus d'un réservoir. Plusieurs modèles initiaux et stratégies d'inversion sont testés afin de minimiser le problème de saut de phase, tout en construisant des modèles de sous-sol avec une résolution suffisante. Sous réserve de mettre en œuvre des stratégies limitant le problème de saut de phase, la JFWI montre qu'elle peut produire un modèle de vitesse acceptable, injectant les bas nombres d'onde dans le modèle de vitesse. L'amélioration de l'éclairage en angles de diffraction fournie par des acquisitions 3D devrait permettre de pouvoir commencer l'inversion par JFWI à partir de modèle encore moins bien définis. / Full waveform inversion (FWI) has attracted worldwide interest for its capacity to estimate the physical properties of the subsurface in details. It is often formulated as a least-squares data-fitting procedure and routinely solved by linearized optimization methods. However, FWI is well known to suffer from cycle skipping problem making the final estimations strongly depend on the user-defined initial models. Reflection waveform inversion (RWI) is recently proposed to mitigate such cycle skipping problem by assuming a scale separation between the background velocity and high-wavenumber reflectivity. It explicitly considers reflected waves such that large-wavelength variations of deep zones can be extracted at the early stage of inversion. Yet, the large-wavelength information of the near surface carried by transmitted waves is neglected.In this thesis, the sensitivity of FWI and RWI to subsurface wavenumbers is revisited in the frame of diffraction tomography and orthogonal decompositions. Based on this analysis, I propose a new method, namely joint full waveform inversion (JFWI), which combines the transmission-oriented FWI and RWI in a unified formulation for a joint sensitivity to low wavenumbers from wide-angle arrivals and short-spread reflections. High-wavenumber components are naturally attenuated during the computation of model updates. To meet the scale separation assumption, I also use a subsurface parameterization based on compressional velocity and acoustic impedance. The temporal complexity of this approach is twice of FWI and the memory requirement is the same.An integrated workflow is then proposed to build the subsurface velocity and impedance models in an alternate way by JFWI and waveform inversion of the reflection data, respectively. In the synthetic example, JFWI is applied to a streamer seismic data set computed in the synthetic Valhall model, the large-wavelength characteristics of which are missing in the initial 1D model. While FWI converges to a local minimum, JFWI succeeds in building a reliable velocity macromodel. Compared with RWI, the involvement of diving waves in JFWI improves the reconstruction of shallow velocities, which translates into an improved imaging at greater depths. The smooth velocity model built by JFWI can be subsequently taken as the initial model for conventional FWI to inject high-wavenumber content without obvious cycle skipping problems.The main promises and limitations of the approach are also reviewed in the real-data application on the 2D OBC profile cross-cutting gas cloud.Several initial models and offset-driven strategies are tested with the aim to manage cycle skipping while building subsurface models with sufficient resolution. JFWI can produce an acceptable velocity model provided that the cycle skipping problem is mitigated and sufficient low-wavenumber content is recovered at the early stage of inversion. Improved scattering-angle illumination provided by 3D acquisitions would allow me to start from cruder initial models.

Le radar géologique est une méthode d'investigation géophysique basée sur la propagation d'ondes électromagnétiques dans le sous-sol. Avec des fréquences allant de 5 MHz à quelques GHz et une forte sensibilité aux propriétés électriques, le géoradar fournit des images de réflectivité dans des contextes et à des échelles très variés : génie civil, géologie, hydrogéologie, glaciologie, archéologie. Cependant, dans certains cas, la compréhension fine des processus étudiés dans la subsurface nécessite une quantification des paramètres physiques du sous-sol. Dans ce but, l'inversion des formes d'ondes complètes, méthode initialement développée pour l'exploration sismique qui exploite l'ensemble des signaux enregistrés, pourrait s'avérer efficace. Dans cette thèse, je propose ainsi des développements méthodologiques par une approche d'inversion multiparamètres (permittivité diélectrique et conductivité), pour des configurations en transmission, en deux dimensions.Ces développements sont ensuite appliqués à un jeu de données réelles acquises entre forages.Dans une première partie, je présente tout d'abord la méthode numérique utilisée pour modéliser la propagation des ondes électromagnétiques dans un milieu 2D hétérogène, élément indispensable pour mener à bien le processus d'imagerie. Ensuite, j’introduis puis étudie le potentiel des méthodes d’optimisation locale standards (gradient conjugué non linéaire, l-BFGS, Newton tronqué dans ses versions Gauss-Newton et Exact-Newton) pour découpler la permittivité diélectrique et la conductivité électrique. Je montre notamment qu’un découplage effectif n’est possible qu’avec un modèle initial suffisamment précis et la méthode la plus sophistiquée (Newton tronqué). Comme dans le cas général, ce modèle initial n’est pas disponible, il s’avère nécessaire d'introduire un facteur d'échelle qui répartit le poids relatif de chaque classe de paramètres dans l'inversion. Dans un milieu réaliste avec une acquisition entre puits, je montre que les différentes méthodes d'optimisation donnent des résultats similaires en matière de découplage de paramètres. C'est finalement la méthode l-BFGS qui est retenue pour l'application aux données réelles, en raison de coûts de calcul plus faibles.Dans une deuxième partie, j'applique cette méthodologie à des données réelles acquises entre deux forages localisés dans des formations carbonatées, à Rustrel (France, 84). Cette inversion est réalisée en parallèle d'une approche synthétique à l'aide d'un modèle représentatif du site étudié et des configurations d'acquisition similaires. Ceci permet de pouvoir comprendre, contrôler et valider les observations et conclusions obtenues sur les données réelles. Cette démarche montre que la reconstruction de la permittivité est très robuste. A contrario, l'estimation de la conductivité souffre de deux couplages majeurs, avec la permittivité diélectrique, d'une part, et avec l'amplitude de la source estimée, d'autre part. Les résultats obtenus sont confrontés avec succès à des données indépendantes (géophysique depuis la surface, analyse sur échantillons de roche), et permet de bénéficier d'une image haute-résolution des formations géologiques. Enfin, une analyse 3D confirme que les structures 3D à fort contraste de propriétés, telles que la galerie enfouie sur notre site, nécessiteraient une approche de modélisation 3D, notamment pour mieux expliquer les amplitudes observées. / Ground Penetrating Radar (GPR) is a geophysical investigation method based on electromagnetic waves propagation in the underground. With frequencies ranging from 5 MHz to a few GHz and a high sensitivity to electrical properties, GPR provides reflectivity images in a wide variety of contexts and scales: civil engineering, geology, hydrogeology, glaciology, archeology. However, in some cases, a better understanding of some subsurface processes requires a quantification of the physical parameters of the subsoil. For this purpose, inversion of full waveforms, a method initially developed for seismic exploration that exploits all the recorded signals, could prove effective. In this thesis, I propose methodological developments using a multiparameter inversion approach (dielectric permittivity and conductivity), for two-dimensional transmission configurations. These developments are then applied to a real data set acquired between boreholes.In a first part, I present the numerical method used to model the propagation of electromagnetic waves in a heterogeneous 2D environment, a much-needed element to carry out the process of imaging. Then, I introduce and study the potential of standard local optimization methods (nonlinear conjugate gradient, l-BFGS, Newton truncated in its Gauss-Newton and Exact-Newton versions) to fight the trade-off effects related to the dielectric permittivity and to the electrical conductivity. In particular, I show that effective decoupling is possible only with a sufficiently accurate initial model and the most sophisticated method (truncated Newton). As in the general case, this initial model is not available, it is necessary to introduce a scaling factor which distributes the relative weight of each parameter class in the inversion. In a realistic medium and for a cross-hole acquisition configuration, I show that the different optimization methods give similar results in terms of parameters decoupling. It is eventually the l-BFGS method that is used for the application to the real data, because of lower computation costs.In a second part, I applied the developed Full waveform inversion methodology to a set of real data acquired between two boreholes located in carbonate formations, in Rustrel (France, 84). This inversion is carried out together with a synthetic approach using a model representative of the studied site and with a similar acquisition configuration. This approach enables us to monitor and validate the observations and conclusions derived from data inversion. It shows that reconstruction of dielectrical permittivity is very robust. Conversely, conductivity estimation suffers from two major couplings: the permittivity and the amplitude of the estimated source. The derived results are successfully compared with independent data (surface geophysics and rock analysis on plugs) and provides a high resolution image of the geological formation. On the other hand, a 3D analysis confirms that 3D structures presenting high properties contrasts, such as the buried gallery present in our site, would require a 3D approach, notably to better explain the observed amplitudes.

We discuss a full-waveform based material profile reconstruction in two-dimensional heterogeneous semi-infinite domains. In particular, we try to image the spatial variation of shear moduli/wave velocities, directly in the time-domain, from scant surficial measurements of the domain's response to prescribed dynamic excitation. In addition, in one-dimensional media, we try to image the spatial variability of elastic and attenuation properties simultaneously. To deal with the semi-infinite extent of the physical domains, we introduce truncation boundaries, and adopt perfectly-matched-layers (PMLs) as the boundary wave absorbers. Within this framework we develop a new mixed displacement-stress (or stress memory) finite element formulation based on unsplit-field PMLs for transient scalar wave simulations in heterogeneous semi-infinite domains. We use, as is typically done, complex-coordinate stretching transformations in the frequency-domain, and recover the governing PDEs in the time-domain through the inverse Fourier transform. Upon spatial discretization, the resulting equations lead to a mixed semi-discrete form, where both displacements and stresses (or stress histories/memories) are treated as independent unknowns. We propose approximant pairs, which numerically, are shown to be stable. The resulting mixed finite element scheme is relatively simple and straightforward to implement, when compared against split-field PML techniques. It also bypasses the need for complicated time integration schemes that arise when recent displacement-based formulations are used. We report numerical results for 1D and 2D scalar wave propagation in semi-infinite domains truncated by PMLs. We also conduct parametric studies and report on the effect the various PML parameter choices have on the simulation error. To tackle the inversion, we adopt a PDE-constrained optimization approach, that formally leads to a classic KKT (Karush-Kuhn-Tucker) system comprising an initial-value state, a final-value adjoint, and a time-invariant control problem. We iteratively update the velocity profile by solving the KKT system via a reduced space approach. To narrow the feasibility space and alleviate the inherent solution multiplicity of the inverse problem, Tikhonov and Total Variation (TV) regularization schemes are used, endowed with a regularization factor continuation algorithm. We use a source frequency continuation scheme to make successive iterates remain within the basin of attraction of the global minimum. We also limit the total observation time to optimally account for the domain's heterogeneity during inversion iterations. We report on both one- and two-dimensional examples, including the Marmousi benchmark problem, that lead efficiently to the reconstruction of heterogeneous profiles involving both horizontal and inclined layers, as well as of inclusions within layered systems. / text