ADArnab Dhara

Arnab Dhara

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Arnab Dhara has a total of 8 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Physics‐Constrained Variational Autoencoder for Uncertainty Quantification of Full Waveform Inversion

    by Abdelrahman Elmeliegy, Mrinal K. Sen, Arnab Dhara, Jennifer Harding, Hongkyu Yoon

    We propose a variational autoencoder framework to directly assess uncertainties in subsurface models produced by single‐ and multiparameter full waveform inversion (FWI). The new method does not require pretraining on labeled data, thus it significantly reduces computational cost and storage requirements. The framework takes the seismic shot gathers as input and returns a set of possible velocity models that fit the input data. The network has three key components: (a) an encoder that maps the input seismic shot gathers to feature distributions in the latent space, (b) a latent vector sampled directly from feature distributions, and (c) a decoder that maps the sampled latent vector to the velocity model space. We incorporated Deep Image Prior principles, leveraging convolutional layers and LeakyReLU activations to regularize the inversion and improve reconstruction. At each epoch, the reconstructed velocity model is passed to a finite difference partial differential equation solver for forward modeling. We then calculate the data misfit between the input seismic data and the simulated data. Additionally, we compute the FWI gradient by cross‐correlating the adjoint and forward wavefield; the FWI gradient is then back‐propagated to guide the update of the neural networks weights and biases. The injection of physics‐based gradients further constrains the inversion and enhancing convergence. We tested the framework on synthetic acoustic and elastic seismic data sets, inverting models of varying complexity and quantifying their associated uncertainties. The results show accurate mean velocity models with meaningful uncertainty estimates, highlighting the potential of the proposed method for practical FWI applications.

  • Elastic Full-Waveform Inversion Using a Physics-Guided Deep Convolutional Encoder–Decoder

    by Arnab Dhara, Mrinal K. Sen

    Elastic full waveform inversion can construct high-resolution P-wave, S-wave velocity and Density models in complex geological settings. However, several factors make the application of elastic FWI challenging. Elastic FWI is prone to the problem of cycle skipping phenomenon when low-frequency in the data are unavailable and the starting model is inaccurate. Multiparameter FWI also suffers from crosstalk issues due to coupling between different model parameters. We extend our physics guided deep convolutional encoder-decoder network to the problem of multiparameter elastic full waveform inversion. Our training is completely unsupervised. Our encoder-decoder which is composed of convolutional neural networks (CNNs) maps the multicomponent shot gathers to the target velocity models. The output from the network is given as input to partial differential equations which generate synthetic data. We compare the observed data against the synthetic data and then compute the misfit. We calculate the gradient of the misfit with respect to the model parameters and then use it to update the neural network weights. We note that the neural network generates velocity and density models that explain the observed data. A toy model, the marmousi model and left part of the BP salt model are used to demonstrate the effectiveness of the proposed approach. Finally, we explain the proposed approach’s efficacy by examining the nature of the loss landscape of neural networks based full waveform inversion.


  • Facies‐constrained transdimensional amplitude versus angle inversion using machine learning assisted priors

    by Arnab Dhara, Mrinal K. Sen, Reetam Biswas

    We present a methodology for seismic inversion that generates high‐resolution models of facies and elastic properties from pre‐stack data. Our inversion algorithm uses a transdimensional approach where, in addition to the layer properties, the number of layers is treated as unknown. In other words, the data itself determine the correct model parameterization, that is, the number of layers. The reversible jump Markov Chain Monte Carlo method is an effective tool to solve such transdimensional problems as it generates models of reservoir properties along with uncertainty estimates. However, current implementations of the reversible jump Markov Chain Monte Carlo algorithms do not account for the non‐Gaussian and multimodal nature of model parameters. The target elastic reservoir properties generally have multimodal and non‐parametric distribution at each location of the model. The number of modes is equal to the number of facies. Taking these factors into account, we extend the reversible jump Markov Chain Monte Carlo algorithm to simultaneously invert for discrete facies and continuous elastic reservoir properties. The proposed extension to the algorithm iteratively samples the facies, by moving from one mode to another, and elastic properties by sampling within the same mode. The integration of facies classification within the inversion reduces non‐uniqueness, improves convergence speed and produces geologically consistent results. The workflow uses machine learning to generate probabilistic priors for the model parameters. We validate our approach by applying it to a synthetic dataset generated from a well log with two facies and then to a complex synthetic two‐dimensional model involving three facies having overlapping elastic property distribution. Finally, we apply our algorithm to a field dataset acquired over an unconventional reservoir. Our algorithm demonstrates the usefulness of incorporating facies information in seismic inversion and also the feasibility of inverting for facies from seismic data.


  • Open Access

    Elastic Full Waveform Inversion using a Physics guided deep convolutional autoencoder

    by Arnab Dhara, Mrinal K. Sen

    Elastic full waveform inversion can construct highresolution P-wave and S-wave velocity models in complex geological settings. However, several factors make the application of elastic FWI challenging. Elastic FWI is prone to the problem of cycle skipping phenomenon when low-frequency in the data are unavailable and the starting model is inaccurate. Multiparameter FWI also suffers from crosstalk issue due to coupling between different model parameters. We extend our physics guided deep convolutional autoencoder network to the problem of multiparameter elastic full waveform inversion. Our training is completely unsupervised. Our autoencoder which is composed of convolutional neural networks (CNNs) maps the multicomponent shot gathers to the target velocity models. The output from the network is given as input to partial differential equations which generate synthetic data. We compare the observed data against the synthetic data and then compute the misfit. We calculate the gradient of the misfit with respect to the model parameters and then use it to update the neural network weights. We note that the neural network generates velocity models that explains the observed data. We demonstrate that the network can introduces regularization in the inversion and overcome issues related to cycle skipping and parameter crosstalks. A toy model, the marmousi model and left part of the BP salt model are used to demonstrate the effectiveness of the proposed approach. Finally, we provide an explanation of the efficacy of the proposed approach by examining the nature of the loss landscape of neural networks based full waveform inversion.

  • Physics-guided deep autoencoder to overcome the need for a starting model in full-waveform inversion

    by Arnab Dhara, Mrinal K. Sen

    Full-waveform inversion (FWI) is a popular technique to obtain high-resolution estimates of earth model parameters using all information present in seismic data. Thus, it can provide important information about the subsurface. The FWI algorithm is formulated as a data-fitting minimization problem that iteratively updates an initial velocity model using the gradient of the misfit until an acceptable match is obtained between the real and synthetic data under a tolerance level based on noise in the data. The inversion is computationally expensive and can converge to a local minimum if the starting model used is not close enough to an optimal model. Here, we propose an alternative approach using a combination of machine learning and the physics of the forward model. Unlike conventional supervised machine learning, known answers are not required to train our network. The shot gathers are input to a convolutional neural network-based autoencoder, the output of which is used as the velocity model that is used to compute synthetic seismograms. The synthetic data are compared against observed input data, and the misfit is estimated. The gradient of the misfit with respect to the velocity model parameters is calculated using the adjoint state method. The adjoint state gradient is then used to update the network weights using the automatic differentiation technique. Once the misfit term converges, the neural network can generate velocity models consistent with the observed data. We observe that the neural network can capture spatial correlations at different scales and thus can introduce regularization in our inverse problem. Experiments with the Marmousi model and SEG Advanced Modeling Corporation Phase 1 salt model suggest that the proposed method can overcome local minima, requires no starting model, and produces robust results in the presence of noise and complex salt body structures.


  • Seismic image registration using multiscale convolutional neural networks

    by Arnab Dhara, Claudio Bagaini

    Aligning seismic images is important in many areas of seismic processing such as time-lapse studies, tomography, and registration of compressional and shear-wave images. This problem is especially difficult when the misalignment is large and varies rapidly and when the images are not shifted versions of each other because they are either contaminated by noise or have different phase or frequency content. In addition, the images may be related by multidimensional vector-valued shift functions. We have developed a fast, scalable, and end-to-end trainable convolutional neural network (CNN) for seismic image registration. The concept of optical flow is widely applied to the problem of image registration using variational methods. Recent developments in the field of computer vision have shown that optical flow estimation can be formulated as a supervised machine learning task and can be successfully solved using CNNs. We train our CNN, SeisFlowNet, on images warped with known shifts and corrupted with noise, frequency, and phase perturbations. We evaluate the promising performance of the trained SeisFlowNet with synthetic data sets where the shift function is known and the images are contaminated with noise and other perturbations. The accuracy of the results obtained with SeisFlowNet is favorably compared with two other popular methods for seismic registration: windowed crosscorrelation and dynamic image warping. Further, we highlight the principles adopted to create training data sets and the advantages and disadvantages of the method.