arXiv subjects

Tariq Alkhalifah

Publications and source records attributed to Tariq Alkhalifah.

At least 91 records · Page 5Linked to original sources

arXiv2022

Laplace HypoPINN: Physics-Informed Neural Network for hypocenter localization and its predictive uncertainty

Several techniques have been proposed over the years for automatic hypocenter localization. While those techniques have pros and cons that trade-off computational efficiency and the susceptibility of getting trapped in local minima, an alternate approach is needed that allows robust localization performance and holds the potential to make the elusive goal of real-time microseismic monitoring possible. Physics-informed neural networks (PINNs) have appeared on the scene as a flexible and versatile framework for solving partial differential equations (PDEs) along with the associated initial or boundary conditions. We develop HypoPINN -- a PINN-based inversion framework for hypocenter localization and introduce an approximate Bayesian framework for estimating its predictive uncertainties. This work focuses on predicting the hypocenter locations using HypoPINN and investigates the propagation of uncertainties from the random realizations of HypoPINN's weights and biases using the Laplace approximation. We train HypoPINN to obtain the optimized weights for predicting hypocenter location. Next, we approximate the covariance matrix at the optimized HypoPINN's weights for posterior sampling with the Laplace approximation. The posterior samples represent various realizations of HypoPINN's weights. Finally, we predict the locations of the hypocenter associated with those weights' realizations to investigate the uncertainty propagation that comes from those realisations. We demonstrate the features of this methodology through several numerical examples, including using the Otway velocity model based on the Otway project in Australia.

cs.LG↗

arXiv2022

$ω$-FWI: Robust full-waveform inversion with Fourier-based metric

Full-waveform inversion is a cutting-edge methodology for recovering high-resolution subsurface models. However, one of the main conventional full-waveform optimization problems challenges is cycle-skipping, usually leading us to an inaccurate local minimum model. A highly investigated track to alleviate this challenge involves designing a more global measure of misfit between the observed and modelled data beyond the sample-to-sample comparison. However, most of these approaches admit relatively smooth inversion results. Here, we introduce a novel misfit function based on the Fourier-based metric. This metric has been successfully applied in molecular physics for solving the Boltzmann equation, and we adapt it to full-waveform inversion. This misfit function exploits the power spectrum information between the modelled and observed data to provide low-wavenumber velocity model updates early, and more high resolution updates as we approach the solution. Thus, it also can be reformulated as a weighted $\ell_{2}$-norm in a quadratic case, which can be seen as a simple extension for conventional full-waveform inversion. Thus, despite its robustness to cycle skipping, it is capable of delivering high-resolution models synonymous to conventional FWI. Considering its frequency domain utilization, we refer to this inversion method as $ω$-FWI. Through the synthetic Marmousi model example, this method successfully recovers an accurate velocity model, starting from a linearly increasing model even for the case of noisy observed data and the lack of low frequencies below 3 Hz and 5Hz, in which the conventional $\ell_{2}$-norm full-waveform inversion suffers from cycle skipping.

physics.comp-ph↗

arXiv2022

StorSeismic: A new paradigm in deep learning for seismic processing

Machine learned tasks on seismic data are often trained sequentially and separately, even though they utilize the same features (i.e. geometrical) of the data. We present StorSeismic, as a framework for seismic data processing, which consists of neural network pre-training and fine-tuning procedures. We, specifically, utilize a neural network as a preprocessing model to store seismic data features of a particular dataset for any downstream tasks. After pre-training, the resulting model can be utilized later, through a fine-tuning procedure, to perform tasks using limited additional training. Used often in Natural Language Processing (NLP) and lately in vision tasks, BERT (Bidirectional Encoder Representations from Transformer), a form of a Transformer model, provides an optimal platform for this framework. The attention mechanism of BERT, applied here on a sequence of traces within the shot gather, is able to capture and store key geometrical features of the seismic data. We pre-train StorSeismic on field data, along with synthetically generated ones, in the self-supervised step. Then, we use the labeled synthetic data to fine-tune the pre-trained network in a supervised fashion to perform various seismic processing tasks, like denoising, velocity estimation, first arrival picking, and NMO. Finally, the fine-tuned model is used to obtain satisfactory inference results on the field data.

cs.LG↗

arXiv2022

Time-lapse data matching using a recurrent neural network approach

Time-lapse seismic data acquisition is an essential tool to monitor changes in a reservoir due to fluid injection, such as CO$_2$ injection. By acquiring multiple seismic surveys in the exact location, we can identify the reservoir changes by analyzing the difference in the data. However, such analysis can be skewed by the near-surface seasonal velocity variations, inaccuracy in the acquisition parameters, and other inevitable noise. The common practice (cross-equalization) to address this problem uses the part of the data where changes are not expected to design a matching filter and then apply it to the whole data, including the reservoir area. Like cross-equalization, we train a recurrent neural network on parts of the data excluding the reservoir area and then infer the reservoir-related data. The recurrent neural network can learn the time dependency of the data, unlike the matching filter that processes the data based on the local information obtained in the filter window. We demonstrate the method of matching the data in various examples and compare it with the conventional matching filter. Specifically, we start by demonstrating the ability of the approach in matching two traces and then test the method on a pre-stack 2D synthetic data. Then, we verify the enhancements of the 4D signal by providing RTM images. We measure the repeatability using normalized root mean square and predictability metrics and show that in some cases, it overcomes the matching filter approach.

physics.geo-ph↗

arXiv2022

Deep learning unflooding for robust subsalt waveform inversion

Full-waveform inversion (FWI), a popular technique that promises high-resolution models, has helped in improving the salt definition in inverted velocity models. The success of the inversion relies heavily on having prior knowledge of the salt, and using advanced acquisition technology with long offsets and low frequencies. Salt bodies are often constructed by recursively picking the top and bottom of the salt from seismic images corresponding to tomography models, combined with flooding techniques. The process is time-consuming and highly prone to error, especially in picking the bottom of the salt (BoS). Many studies suggest performing FWI with long offsets and low frequencies after constructing the salt bodies to correct the miss-interpreted boundaries. Here, we focus on detecting the BoS automatically by utilizing deep learning tools. We specifically generate many random 1D models, containing or free of salt bodies, and calculate the corresponding shot gathers. We then apply FWI starting with salt flooded versions of those models, and the results of the FWI become inputs to the neural network, whereas the corresponding true 1D models are the output. The network is trained in a regression manner to detect the BoS and estimate the subsalt velocity. We analyze three scenarios in creating the training datasets and test their performance on the 2D BP 2004 salt model. We show that when the network succeeds in estimating the subsalt velocity, the requirement of low frequencies and long offsets are somewhat mitigated. In general, this work allows us to merge the top-to-bottom approach with FWI, save the BoS picking time, and empower FWI to converge in the absence of low frequencies and long offsets in the data.

physics.geo-ph↗

arXiv2021

PINNup: Robust neural network wavefield solutions using frequency upscaling and neuron splitting

Solving for the frequency-domain scattered wavefield via physics-informed neural network (PINN) has great potential in seismic modeling and inversion. However, when dealing with high-frequency wavefields, its accuracy and training cost limits its applications. Thus, we propose a novel implementation of PINN using frequency upscaling and neuron splitting, which allows the neural network model to grow in size as we increase the frequency while leveraging the information from the pre-trained model for lower-frequency wavefields, resulting in fast convergence to high-accuracy solutions. Numerical results show that, compared to the commonly used PINN with random initialization, the proposed PINN exhibits notable superiority in terms of convergence and accuracy and can achieve neuron based high-frequency wavefield solutions with a two-hidden-layer model.

cs.LG↗

arXiv2021

The potential of self-supervised networks for random noise suppression in seismic data

Noise suppression is an essential step in any seismic processing workflow. A portion of this noise, particularly in land datasets, presents itself as random noise. In recent years, neural networks have been successfully used to denoise seismic data in a supervised fashion. However, supervised learning always comes with the often unachievable requirement of having noisy-clean data pairs for training. Using blind-spot networks, we redefine the denoising task as a self-supervised procedure where the network uses the surrounding noisy samples to estimate the noise-free value of a central sample. Based on the assumption that noise is statistically independent between samples, the network struggles to predict the noise component of the sample due to its randomnicity, whilst the signal component is accurately predicted due to its spatio-temporal coherency. Illustrated on synthetic examples, the blind-spot network is shown to be an efficient denoiser of seismic data contaminated by random noise with minimal damage to the signal; therefore, providing improvements in both the image domain and down-the-line tasks, such as inversion. To conclude the study, the suggested approach is applied to field data and the results are compared with two commonly used random denoising techniques: FX-deconvolution and Curvelet transform. By demonstrating that blind-spot networks are an efficient suppressor of random noise, we believe this is just the beginning of utilising self-supervised learning in seismic applications.

physics.geo-ph↗

arXiv2021

MLReal: Bridging the gap between training on synthetic data and real data applications in machine learning

Among the biggest challenges we face in utilizing neural networks trained on waveform data (i.e., seismic, electromagnetic, or ultrasound) is its application to real data. The requirement for accurate labels forces us to develop solutions using synthetic data, where labels are readily available. However, synthetic data often do not capture the reality of the field/real experiment, and we end up with poor performance of the trained neural network (NN) at the inference stage. We describe a novel approach to enhance supervised training on synthetic data with real data features (domain adaptation). Specifically, for tasks in which the absolute values of the vertical axis (time or depth) of the input data are not crucial, like classification, or can be corrected afterward, like velocity model building using a well-log, we suggest a series of linear operations on the input so the training and application data have similar distributions. This is accomplished by applying two operations on the input data to the NN model: 1) The crosscorrelation of the input data (i.e., shot gather, seismic image, etc.) with a fixed reference trace from the same dataset. 2) The convolution of the resulting data with the mean (or a random sample) of the autocorrelated data from another domain. In the training stage, the input data are from the synthetic domain and the auto-correlated data are from the real domain, and random samples from real data are drawn at every training epoch. In the inference/application stage, the input data are from the real subset domain and the mean of the autocorrelated sections are from the synthetic data subset domain. Example applications on passive seismic data for microseismic event source location determination and active seismic data for predicting low frequencies are used to demonstrate the power of this approach in improving the applicability of trained models to real data.

physics.geo-ph↗

arXiv2021

Direct domain adaptation through reciprocal linear transformations

We propose a direct domain adaptation (DDA) approach to enrich the training of supervised neural networks on synthetic data by features from real-world data. The process involves a series of linear operations on the input features to the NN model, whether they are from the source or target domains, as follows: 1) A cross-correlation of the input data (i.e. images) with a randomly picked sample pixel (or pixels) of all images from that domain or the mean of all randomly picked sample pixel (or pixels) of all images. 2) The convolution of the resulting data with the mean of the autocorrelated input images from the other domain. In the training stage, as expected, the input images are from the source domain, and the mean of auto-correlated images are evaluated from the target domain. In the inference/application stage, the input images are from the target domain, and the mean of auto-correlated images are evaluated from the source domain. The proposed method only manipulates the data from the source and target domains and does not explicitly interfere with the training workflow and network architecture. An application that includes training a convolutional neural network on the MNIST dataset and testing the network on the MNIST-M dataset achieves a 70% accuracy on the test data. A principal component analysis (PCA), as well as t-SNE, show that the input features from the source and target domains, after the proposed direct transformations, share similar properties along with the principal components as compared to the original MNIST and MNIST-M input features.

cs.LG↗

arXiv2021

Solving the Eikonal equation for compressional and shear waves in anisotropic media using peridynamic differential operator

The traveltime of compressional (P) and shear (S) waves have proven essential in many applications of earthquake and exploration seismology. An accurate and efficient traveltime computation for P and S waves is crucial for the success of these applications. However, solutions to the Eikonal equation with a complex phase velocity field in anisotropic media is challenging. The Eikonal equation is a first-order, hyperbolic, nonlinear partial differential equation (PDE) that represents the high-frequency asymptotic approximation of the wave equation. The fast marching and sweeping methods are commonly used due to their efficiency in numercally solving Eikonal equation. However, these methods suffer from numerical inaccuracy in anisotropic media with sharp heterogeneity, irregular surface topography and complex phase velocity fields. This study presents a new method to solving the Eikonal equation by employing the peridynamic differential operator (PDDO). The PDDO provides the nonlocal form of the Eikonal equation by introducing an internal length parameter (horizon) and a weight function with directional nonlocality. The operator is immune to discontinuities in the form sharp changes in field or model variables and invokes the direction of traveltime in a consistent manner. The weight function controls the degree of association among points within the horizon. Solutions are constructed in a consistent manner without upwind assumptions through simple discretization. The capability of this approach is demonstrated by considering different types of Eikonal equations on complex velocity models in anisotropic media. The examples demonstrate its unconditional numerical stability and results compare well with the reference solutions.

physics.comp-ph↗

arXiv2021

Wavefield solutions from machine learned functions

Solving the wave equation is one of the most (if not the most) fundamental problems we face as we try to illuminate the Earth using recorded seismic data. The Helmholtz equation provides wavefield solutions that are dimensionally reduced, per frequency, compared to the time domain, which is useful for many applications, like full waveform inversion (FWI). However, our ability to attain such wavefield solutions depends often on the size of the model and the complexity of the wave equation. Thus, we use here a recently introduced framework based on neural networks to predict functional solutions through setting the underlying physical equation as a loss function to optimize the neural network parameters. For an input given by a location in the model space, the network learns to predict the wavefield value at that location, and its partial derivatives using a concept referred to as automatic differentiation, to fit, in our case, a form of the Helmholtz equation. We specifically seek the solution of the scattered wavefield considering a simple homogeneous background model that allows for analytical solutions of the background wavefield. Providing the neural network (NN) a reasonable number of random points from the model space will ultimately train a fully connected deep NN to predict the scattered wavefield function. The size of the network depends mainly on the complexity of the desired wavefield, with such complexity increasing with increasing frequency and increasing model complexity. However, smaller networks can provide smoother wavefields that might be useful for inversion applications. Preliminary tests on a two-box-shaped scatterer model with a source in the middle, as well as, the Marmousi model with a source on the surface demonstrate the potential of the NN for this application. Additional tests on a 3D model demonstrate the potential versatility of the approach.

physics.geo-ph↗

arXiv2021

A holistic approach to computing first-arrival traveltimes using neural networks

Since the original algorithm by John Vidale in 1988 to numerically solve the isotropic eikonal equation, there has been tremendous progress on the topic addressing an array of challenges including improvement of the solution accuracy, incorporation of surface topography, adding more accurate physics by accounting for anisotropy/attenuation in the medium, and speeding up computations using multiple CPUs and GPUs. Despite these advances, there is no mechanism in these algorithms to carry on information gained by solving one problem to the next. Moreover, these approaches may breakdown for certain complex forms of the eikonal equation, requiring approximation methods to estimate the solution. Therefore, we seek an alternate approach to address the challenge in a holistic manner, i.e., a method that not only makes it simpler to incorporate topography, allow accounting for any level of complexity in physics, benefiting from computational speedup due to the availability of multiple CPUs or GPUs, but also able to transfer knowledge gained from solving one problem to the next. We develop an algorithm based on the emerging paradigm of physics-informed neural network to solve various forms of the eikonal equation. We show how transfer learning and surrogate modeling can be used to speed up computations by utilizing information gained from prior solutions. We also propose a two-stage optimization scheme to expedite the training process in presence of sharper heterogeneity in the velocity model. Furthermore, we demonstrate how the proposed approach makes it simpler to incorporate additional physics and other features in contrast to conventional methods that took years and often decades to make these advances. Such an approach not only makes the implementation of eikonal solvers much simpler but also puts us on a much faster path to progress.

physics.comp-ph↗

arXiv2021

PINNeik: Eikonal solution using physics-informed neural networks

The eikonal equation is utilized across a wide spectrum of science and engineering disciplines. In seismology, it regulates seismic wave traveltimes needed for applications like source localization, imaging, and inversion. Several numerical algorithms have been developed over the years to solve the eikonal equation. However, these methods require considerable modifications to incorporate additional physics, such as anisotropy, and may even breakdown for certain complex forms of the eikonal equation, requiring approximation methods. Moreover, they suffer from computational bottleneck when repeated computations are needed for perturbations in the velocity model and/or the source location, particularly in large 3D models. Here, we propose an algorithm to solve the eikonal equation based on the emerging paradigm of physics-informed neural networks (PINNs). By minimizing a loss function formed by imposing the eikonal equation, we train a neural network to output traveltimes that are consistent with the underlying partial differential equation. We observe sufficiently high traveltime accuracy for most applications of interest. We also demonstrate how the proposed algorithm harnesses machine learning techniques like transfer learning and surrogate modeling to speed up traveltime computations for updated velocity models and source locations. Furthermore, we use a locally adaptive activation function and adaptive weighting of the terms in the loss function to improve convergence rate and solution accuracy. We also show the flexibility of the method in incorporating medium anisotropy and free-surface topography compared to conventional methods that require significant algorithmic modifications. These properties of the proposed PINN eikonal solver are highly desirable in obtaining a flexible and efficient forward modeling engine for seismological applications.

physics.comp-ph↗

arXiv2021

Wavefield reconstruction inversion via physics-informed neural networks

Wavefield reconstruction inversion (WRI) formulates a PDE-constrained optimization problem to reduce cycle skipping in full-waveform inversion (FWI). WRI often requires expensive matrix inversions to reconstruct frequency-domain wavefields. Physics-informed neural network (PINN) uses the underlying physical laws as loss functions to train the neural network (NN), and it has shown its effectiveness in solving the Helmholtz equation and generating Green's functions, specifically for the scattered wavefield. By including a data-constrained term in the loss function, the trained NN can reconstruct a wavefield that simultaneously fits the recorded data and satisfies the Helmholtz equation for a given initial velocity model. Using the predicted wavefields, we rely on a small-size NN to predict the velocity using the reconstructed wavefield. In this velocity prediction NN, spatial coordinates are used as input data to the network and the scattered Helmholtz equation is used to define the loss function. After we train this network, we are able to predict the velocity in the domain of interest. We develop this PINN-based WRI method and demonstrate its potential using a part of the Sigsbee2A model and a modified Marmousi model. The results show that the PINN-based WRI is able to invert for a reasonable velocity with very limited iterations and frequencies, which can be used in a subsequent FWI application.

physics.geo-ph↗

arXiv2021

PINNtomo: Seismic tomography using physics-informed neural networks

Seismic traveltime tomography using transmission data is widely used to image the Earth's interior from global to local scales. In seismic imaging, it is used to obtain velocity models for subsequent depth-migration or full-waveform inversion. In addition, cross-hole tomography has been successfully applied for a variety of applications, including mineral exploration, reservoir monitoring, and CO2 injection and sequestration. Conventional tomography techniques suffer from a number of limitations, including the use of a smoothing regularizer that is agnostic to the physics of wave propagation. Here, we propose a novel tomography method to address these challenges using developments in the field of scientific machine learning. Using seismic traveltimes observed at seismic stations covering part of the computational model, we train neural networks to approximate the traveltime factor and the velocity fields, subject to the physics-informed regularizer formed by the factored eikonal equation. This allows us to better compensate for the ill-posedness of the tomography problem compared to conventional methods and results in a number of other attractive features, including computational efficiency. We show the efficacy of the proposed method and its capabilities through synthetic tests for surface seismic and cross-hole geometries. Contrary to conventional techniques, we find the performance of the proposed method to be agnostic to the choice of the initial velocity model.

physics.comp-ph↗

arXiv2020

Solving the acoustic VTI wave equation using physics-informed neural networks

Frequency-domain wavefield solutions corresponding to the anisotropic acoustic wave equations can be used to describe the anisotropic nature of the earth. To solve a frequency-domain wave equation, we often need to invert the impedance matrix. This results in a dramatic increase in computational cost as the model size increases. It is even a bigger challenge for anisotropic media, where the impedance matrix is far more complex. To address this issue, we use the emerging paradigm of physics-informed neural networks (PINNs) to obtain wavefield solutions for an acoustic wave equation for transversely isotropic (TI) media with a vertical axis of symmetry (VTI). PINNs utilize the concept of automatic differentiation to calculate its partial derivatives. Thus, we use the wave equation as a loss function to train a neural network to provide functional solutions to form of the acoustic VTI wave equation. Instead of predicting the pressure wavefields directly, we solve for the scattered pressure wavefields to avoid dealing with the point source singularity. We use the spatial coordinates as input data to the network, which outputs the real and imaginary parts of the scattered wavefields and auxiliary function. After training a deep neural network (NN), we can evaluate the wavefield at any point in space instantly using this trained NN. We demonstrate these features on a simple anomaly model and a layered model. Additional tests on a modified 3D Overthrust model and a model with irregular topography also show the effectiveness of the proposed method.

physics.comp-ph↗

arXiv2020

ML-misfit: Learning a robust misfit function for full-waveform inversion using machine learning

Most of the available advanced misfit functions for full waveform inversion (FWI) are hand-crafted, and the performance of those misfit functions is data-dependent. Thus, we propose to learn a misfit function for FWI, entitled ML-misfit, based on machine learning. Inspired by the optimal transport of the matching filter misfit, we design a neural network (NN) architecture for the misfit function in a form similar to comparing the mean and variance for two distributions. To guarantee the resulting learned misfit is a metric, we accommodate the symmetry of the misfit with respect to its input and a Hinge loss regularization term in a meta-loss function to satisfy the "triangle inequality" rule. In the framework of meta-learning, we train the network by running FWI to invert for randomly generated velocity models and update the parameters of the NN by minimizing the meta-loss, which is defined as accumulated difference between the true and inverted models. We first illustrate the basic principle of the ML-misfit for learning a convex misfit function for travel-time shifted signals. Further, we train the NN on 2D horizontally layered models, and we demonstrate the effectiveness and robustness of the learned ML-misfit by applying it to the well-known Marmousi model.

physics.geo-ph↗

arXiv2020

A data-driven choice of misfit function for FWI using reinforcement learning

In the workflow of Full-Waveform Inversion (FWI), we often tune the parameters of the inversion to help us avoid cycle skipping and obtain high resolution models. For example, typically start by using objective functions that avoid cycle skipping, like tomographic and image based or using only low frequency, and then later, we utilize the least squares misfit to admit high resolution information. We also may perform an isotropic (acoustic) inversion to first update the velocity model and then switch to multi-parameter anisotropic (elastic) inversions to fully recover the complex physics. Such hierarchical approaches are common in FWI, and they often depend on our manual intervention based on many factors, and of course, results depend on experience. However, with the large data size often involved in the inversion and the complexity of the process, making optimal choices is difficult even for an experienced practitioner. Thus, as an example, and within the framework of reinforcement learning, we utilize a deep-Q network (DQN) to learn an optimal policy to determine the proper timing to switch between different misfit functions. Specifically, we train the state-action value function (Q) to predict when to use the conventional L2-norm misfit function or the more advanced optimal-transport matching-filter (OTMF) misfit to mitigate the cycle-skipping and obtain high resolution, as well as improve convergence. We use a simple while demonstrative shifted-signal inversion examples to demonstrate the basic principles of the proposed method.

physics.geo-ph↗