Conference Papers
Recent Submissions

Aiding selfsupervised coherent noise suppression by the introduction of signal segmentation using blindspot networks(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Blindspot networks have been shown to be natural noise suppressors under the assumption that noise is unpredictable based on the information fed into the network during training. Trained in a selfsupervised manner, such approaches only utilise the original raw data to determine to remove the noise. In this work, we propose two novel elements for enhancing blindspot denoising: (1) the introduction of a 2class segmentation task to aid the network in identification of interest areas of signals that require particular attention during denoising, and; (2) the introduction of a tracewise noise mask designed to obscure the coherency of noise from being observed by the network. The joint scheme is achieved by introducing a joint loss function to balance between the two deep learning tasks. As such, the final joint scheme is the combination of a selfsupervised, blindspot denoising procedure and a supervised segmentation procedure. We illustrate how the joint scheme can improve the denoising performance of the network, hypothesising that this is due to the introduction of prior information guiding the denoising procedure to areas of focus. Preliminary results from synthetic data contaminated by tracewise noise, show an increase in the structural similarity index from 0.989 to 0.995, when comparing the optimal jointscheme versus the pure denoising procedure. Future work will extend the procedure to field data where rulebased approaches will be used to generate the segmentation labels.

Bayesian RockAVO: Direct petrophysical inversion with hierarchical conditional GANs(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Reservoir characterization is a critical component in any oil and gas, geothermal, and CO2 sequestration project. A fundamental step in the process of characterizing the subsurface is represented by the inversion of petrophysical parameters from seismic data. However, this problem suffers from various uncertainty sources originating from inaccuracies in the measurements, modeling errors, and complex geological processes. Moreover, the nonlinearity of the rockphysics model and Zoeppritz equation that constitute the modelling operator, further complicates the inversion process. In this work, we propose a novel datadriven approach where welllog information is used to obtain optimal basis functions that link bandlimited petrophysical reflectivities to prestack seismic data. Subsequently, the inversion of such bandlimited reflectivities for petrophysical parameters is framed in a Bayesian framework where a generative adversarial network is used to produce a geologically realistic prior distribution. The trained prior distribution is updated using the Stein Variational Gradient Descent and a set of representative solutions is produced that is consistent with the uncertainties in the data and the nonlinear operators.

Using deep learning for automatic detection and segmentation of carbonate microtextures(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]The difficulties involved in studying micrometersized micrite crystals, and quantifying the associated impact on large scale geophysical properties, have long hindered our society’s understanding of both Middle Eastern and global microporous limestones. Instance segmentation procedures, from the field of deep learning, offer the ability to identify at a pixellevel each individual crystal within an SEM image, allowing for automated morphological analysis. We illustrate how the common Masked Regionbased Convolution Neural Network from computer vision can be adapted to the task of identifying individual micrite crystal within grayscale SEM images. Leveraging Transfer Learning, the ResNet50 neural architecture is used with weights initialized through a pretraining on Microsoft’s Common Objects in COntext (COCO) dataset. The resulting model accurately detects and separates a number of crystals observed within different SEM images. However the trained model is also shown to be highly susceptible to noise introduced as part of the imaging procedure, for example charging noise. Future work will aim to make the procedure more robust, reducing the impact of noise by adapting the preprocessing workflow and incorporating more noisy images into the training dataset.

Pwdpinn: Slopeassisted seismic interpolation with physicsinformed neural networks(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Seismic data can be expressed as a superposition of local plane waves. A gather of traces can be described by the local plane wave differential equation (PDE), that allows to predict each of the traces from the previous one, given the knowledge of the local slope of the events. In the approach presented here, we train a neural network in an unsupervised manner to solve seismic interpolation problems using the local plane wave differential equation and the local slope estimated by the mean of plane wave destruction filters (PWD). The physicsinformed neural network (PINN) maps the input grid points in time and space to the amplitudes of the wavefield whilst matching the information contained in the available traces. The proposed approach is tested on two seismic interpolation tasks using synthetic data, specifically, interpolation of data with large gaps and those aliased. Whilst the network shows remarkable interpolation capabilities in both experiments, it tends to struggle fitting aliased data with high frequency content. To mitigate this problem, we propose to include locally adaptive activation functions in the architecture. This leads to improved convergence and reconstruction accuracy.

Traveltime tomography and efficient physicsinformed Bayesian inversion(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Bayesian inversion provides the same information as regularized inversion of seismic data, except it also supplies a probability estimate of the solution throughout model space. The cost, however, is that Bayesian inversion is ordersofmagnitude more expensive than regularized inversion by a gradient optimization method. To mitigate this cost, we present an efficient physicsinformed Bayesian inversion method that combines regularized inversion to get both the optimal solution and the posterior probability functions in model space. A gradient optimization method is used to efficiently estimate the maximum a posterior (MAP) solution, and so function evaluations are only needed around the MAP point in model space. This efficiently provides the posterior probability in that neighbourhood, and therefore avoids the tremendous expense of sampling points throughout the highdimensional model space. We apply this physicsinformed Bayesian inversion to VSP traveltime data. The tomogram is computed with the assistance of an analytic inverse, and the posterior probability estimate is computed with an orderofmagnitude less cost than standard Bayesian analysis. This procedure can also be adapted to refraction traveltime tomography for nearsurface imaging.

Efficient PhysicsInformed Bayesian Inversion of VSP and Hydrofrac Data(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Bayesian inversion provides the same information as regularized inversion of seismic data, except it also supplies a probability estimate of the solution throughout model space. The cost, however, is that Bayesian inversion is ordersofmagnitude more expensive than regularized inversion by a gradient optimization method. To mitigate this cost, we present an efficient physicsinformed Bayesian inversion method that combines regularized inversion to get both the optimal solution and the posterior probability functions in model space. A gradientoptimization method is used to efficiently estimate the maximum a posterior (MAP) solution, and so function evaluations are only needed around the MAP point in model space. This efficiently provides the posterior probability in that neighborhood, and therefore avoids the tremendous expense of sampling points throughout the highdimensional model space. We present two applications of this physicsinformed Bayesian inversion: VSP traveltime inversion and migration of passive seismic data. For 4D monitoring of hydrofrac operations, reuse of the previously computed traveltimes for probability estimates is ordersofmagnitude less expensive for computing migration and posterior images.

Robust joint inversion and segmentation of 4D seismic data(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Seismic inversion is the leading method to map and quantify changes from timelapse (4D) seismic datasets, with applications ranging from monitoring of hydrocarbonproducing fields to carbon capture and sequestration. Timelapse seismic inversion is however, a notoriously illposed inverse problem: the bandlimited nature of seismic data, alongside inaccuracies in the repeatability of consecutive acquisition surveys make it challenging to obtain highresolution, clean estimates of 4D effects. Adding prior information to the inversion process in the form of properly crafted regularization (or preconditioning) is therefore essential to successfully extract weak signals that are usually buried under strong noise. In this work, we leverage the fact that 4D seismic inversion can be described as a coupled inversion of its baseline and monitor 3D seismic datasets. In existing approaches, the coupling is introduced by penalizing the squared L2norm of difference between the baseline and the monitor acoustic impedances, as this is usually assumed to be small. A major downside of such a regularization is that, whilst reducing the overall level of noise in the estimated acoustic impedance differences, the resulting 4D effects are usually oversmoothed and their strength is underestimated. We instead propose to adapt the joint inversion and segmentation algorithm introduced by Ravasi and Birnie (2021) to the problem of 4D seismic inversion. Our technique produces two acoustic impedance models by inverting the corresponding 3D seismic datasets, regularized by TotalVariation. Moreover, the objective function to optimize is augmented with a segmentation term that renders solutions consistent with the expected 4D effects (obtained, for example, as part of a 4D feasibility study by means of Gassmann fluid substitution). A numerical experiment is presented to validate the effectiveness of the proposed approach and its superiority over stateoftheart 4D inversion methods.

Boosting selfsupervised blindspot networks via transfer learning(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Selfsupervised procedures offer an appealing alternative to supervised denoising techniques that require noisyclean pairs of training data. However, the capabilities of selfsupervised denoising procedures are often limited by the requirement that noise cannot be predicted directly from neighbouring values in the training input samples. As such, there is often a tradeoff with respect to the number of training epochs between learning to replicate the signal without learning to replicate the noise. Focusing on blindspot networks that learn a pixel’s value based on neighbouring pixels, we propose to train a supervised model in a blindspot manner such that the model learns how to predict a pixel’s clean value based off its noisy neighbouring traces. The weights of the trained model are then used to initialise a selfsupervised model which is trained purely on noisy field data. In comparison to the fully selfsupervised approach, we illustrate that pretraining with synthetic data results in increased noise suppression, alongside a lower level of signal leakage in the field data.

Deep Earth: Leveraging neural networks for seismic exploration objectives(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]Machine learning has already made many inroads in developments related to acquisition, processing, imaging, inverting, and interpreting seismic data. In spite of the many success stories, its commercial use has been limited as the challenges mount. These challenges include cost of training, availability of training samples, the applicability of the trained model to real data (generalization), and more importantly, the availability of practitioners who actually know what the neural networks (NNs) are doing. Taking a step back, I will review what worked in deep learning and what we are still waiting on to work. We will look into the various ML algorithms, from supervised to unsupervised, transformers to contrastive learning, and identify the potential role of these various algorithms on seismic data, with examples. The examples include seismic data denoising, data extrapolation, first arrival picking, microseismic location, velocity inversion all on real data.

Largescale Marchenko imaging with distanceaware matrix reordering, tile lowrank compression, and mixedprecision computations(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]A variety of waveequationbased seismic processing algorithms rely on the repeated application of the Multi Dimensional Convolution (MDC) operator. For largescale 3D seismic surveys, this comes with severe computational challenges due to the sheer size of highdensity, fullazimuth seismic datasets required by such algorithms. We present a threefold solution that greatly alleviates the memory footprint and computational cost of 3D MDC by leveraging a combination of i) distanceaware matrix reordering, ii) Tile LowRank (TLR) matrix compression, and iii) computations in mixed floatingpoint precision. By applying our strategy to a 3D synthetic dataset, we show that the size of kernel matrices used in the Marchenko redatuming and MultiDimensional Deconvolution equations can be reduced by a factor of 34 and 6, respectively. We also introduce a TLR MatrixVector Multiplication (TLRMVM) algorithm that, as a direct consequence of such compression capabilities, is consistently faster than its dense counterpart by a factor of 4.8 to 36.1 (depending on the selected hardware). As a result, the associated inverse problems can be solved at a fraction of cost in comparison to stateof theart implementations that require a pass through the entire data at each MDC operation. This is achieved with minimal impact on the quality of the processing outcome.

Physicsbased preconditioned multidimensional deconvolution in the time domain(Society of Exploration Geophysicists and American Association of Petroleum Geologists, 20220815) [Conference Paper]MultiDimensional Deconvolution is a datadriven method that is at the center of key seismic processing applications  from suppressing multiples to inversionbased imaging. When posed in an interferometric context, it can grant access to overburdenfree seismic virtual surveys at a given datum in the subsurface. As such, it constitutes an essential processing operation that achieves multiple imaging objectives simultaneously in redatuming or targetoriented imaging: e.g., suppressing multiples, removing complex overburden effects, and retrieving amplitude consistent image gathers for impedance inversion. Despite its potential, the deconvolution process relies on the solution of an illconditioned linear inverse problem sensitive to noise artifacts due to incomplete acquisition, limited sources, and bandlimited data. Typically, this inversion is performed in the Fourier domain where the estimation of optimal regularization parameters hinders accurate waveform reconstruction. We reformulate the problem in the time domain  long believed to be computationally intractable  and introduce several physical constraints that naturally drive the inversion towards a reduced set of reliable, stable solutions. This allows to successfully reconstruct the overburdenfree reflection response beneath a complex salt body from noisecontaminated data.

DeepLearning Based Channel Estimation for RISAided mmWave Systems with Beam Squint(IEEE, 20220811) [Conference Paper]Reconfigurable intelligent surface (RIS) assisted wireless systems require accurate channel state information (CSI) to control wireless channels and improve overall network performance. However, CSI acquisition is nontrivial due to the passive nature of RIS, and the dimensions of the cascaded channel between transceivers increase with the large number of RIS elements, which requires high training overhead. Prior art has considered frequencyselective channel estimation without considering the beam squint effect in wideband systems, severely degrading channel estimation performance. This paper proposes a novel datadriven approach for estimating wideband cascaded channels of RISassisted multiuser millimeterwave massive multipleinput multipleoutput (MIMO) systems with limited training overhead, explicitly considering the effect of beam squint. To circumvent the beam squint effect, the proposed method exploits the common sparsity property among the different subcarriers as well as the doublestructured sparsity property of the users’ angular cascaded channel matrices. The proposed datadriven cascaded channel estimation approach exploits denoising neural networks to detect channel supports accurately. Compared to beam squint effect agnostic traditional orthogonal matching pursuit (OMP) approaches, the proposed datadriven approach achieves 56dB less normalized mean square error (NMSE) and reduces the lower bound gap to only 1dB for the oracle leastsquare benchmark.

Waveguiding via Transformation Optics(IEEE, 20220809) [Conference Paper]We demonstrate that it is possible to surpass current limitations of nanophotonics and plasmonics by designing an artificial material which can emulate userdefined spatial refractive index distribution. The effective optical property of the material is engineered through the deformation of reflective substrate via transformation optics approach. We provide one of possible applications  subwavelength optical waveguide coupler device based on this technique.

Differentially Private ℓ1norm Linear Regression with Heavytailed Data(IEEE, 20220803) [Conference Paper]We study the problem of Differentially Private Stochastic Convex Optimization (DPSCO) with heavytailed data. Specifically, we focus on the ℓ1norm linear regression in the ϵDP model. While most of the previous work focuses on the case where the loss function is Lipschitz, here we only need to assume the variates has bounded moments. Firstly, we study the case where the ℓ2 norm of data has bounded second order moment. We propose an algorithm which is based on the exponential mechanism and show that it is possible to achieve an upper bound of O~(dnε−−√) (with high probability). Next, we relax the assumption to bounded θth order moment with some θ ∈ (1,2) and show that it is possible to achieve an upper bound of O~((dnε−−√)θ−1θ). Our algorithms can also be extended to more relaxed cases where only each coordinate of the data has bounded moments, and we can get an upper bound of O~(dnε−−√) and O~(d(nε)θ−1θ) in the second and θth moment case respectively.

Setaware Entity Synonym Discovery with Flexible Receptive Fields (Extended Abstract)(IEEE, 20220802) [Conference Paper]Entity synonym discovery (ESD) from text corpus is an essential problem in many entityleveraging applications. This paper aims to address three limitations that widely exist in the current ESD solutions: 1) the lack of effective utilization for synonym set information; 2) the feature extraction of entities from restricted receptive fields; and 3) the incapacity to capture higherorder contextual information. We propose a novel setaware ESD model that enables a flexible receptive field for ESD by using entity synonym set information and constructing a twolevel network. Extensive experimental results on public datasets show that our model consistently outperforms the stateoftheart with significant improvement.

An industrial perspective on web scraping characteristics and open issues(IEEE, 20220725) [Conference Paper]An ongoing battle has been running for more than a decade between ecommerce websites owners and web scrapers. Whenever one party finds a new technique to prevail, the other one comes up with a solution to defeat it. Based on our industrial experience, we know this problem is far from being solved. New solutions are needed to address automated threats. In this work, we will describe the actors taking part in the battle, the weapons at their disposal, and their allies on either side. We will present a realworld setup to explain how ecommerce websites operators try to defend themselves and the open problems they seek solutions for.

CLIP2StyleGAN: Unsupervised Extraction of StyleGAN Edit Directions(ACM, 20220724) [Conference Paper]The success of StyleGAN has enabled unprecedented semantic editing capabilities, on both synthesized and real images. However, such editing operations are either trained with semantic supervision or annotated manually by users. In another development, the CLIP architecture has been trained with internetscale loose image and text pairings, and has been shown to be useful in several zeroshot learning settings. In this work, we investigate how to effectively link the pretrained latent spaces of StyleGAN and CLIP, which in turn allows us to automatically extract semanticallylabeled edit directions from StyleGAN, finding and naming meaningful edit operations, in a fully unsupervised setup, without additional human guidance. Technically, we propose two novel building blocks; one for discovering interesting CLIP directions and one for semantically labeling arbitrary directions in CLIP latent space. The setup does not assume any predetermined labels and hence we do not require any additional supervised text/attributes to build the editing framework. We evaluate the effectiveness of the proposed method and demonstrate that extraction of disentangled labeled StyleGAN edit directions is indeed possible, revealing interesting and nontrivial edit directions.

BeeCast: A DevicetoDevice Collaborative Video Streaming System(IEEE, 20220719) [Conference Paper]In this paper, we propose BeeCast, a collaborative video streaming system that facilitates collaborative video streaming for a group of mobile users with limited Internet connectivity. The novelty of the proposed system is the ability to watch the video on a shared screen or to watch the video on multiple screens. The latter option entails proposing a method to exchange the downloaded video segments among the users using devicetodevice communication. The proposed system is composed of two components: BeeBuzzer, and BeePlanner. The BeeBuzzer component manages and coordinates the segment exchange among devices, while BeePlanner component enhances the overall Quality of Experience (QoE) through effective segment assignments decisions for each user. Simulation results show that using BeeCast in an unstable network produces a more consistent QoE than individual streaming while eliminating 80% of redundant network traffic.

RF Performance Assessment of Sub8nm GaNSOIFinFET Using Power Gain Parameters(IEEE, 20220718) [Conference Paper]In this work, we have presented, a radio frequency (RF) assessment of nanoscale gallium nitridesilicononinsulator fin fieldeffect transistor (GaNSOIFinFET). All the performances of the device have been compared with conventional FinFET (Conv. FinFET) simultaneously. All the results show that the power gains have significantly improved in terms of Gma, Gms, stern stability factor (SS), GMT, and intrinsic delay in comparison to conventional FinFET. Current gain unilateral power gain and have also been evaluated for the extraction of fT (cutoff frequency) and fMAX respectively. fT and fMAX enhance by 88.8% and 94.6% respectively. This analysis has been done at several THz frequencies. The implementation of GaN in the channel reduces the parasitic capacitance and paves the way for highperformance RF applications.

Parallel spacetime likelihood optimization for air pollution prediction on largescale systems(ACM, 20220712) [Conference Paper]Gaussian geostatistical spacetime modeling is an effective tool for performing statistical inference of field data evolving in space and time, generalizing spatial modeling alone at the cost of the greater complexity of operations and storage, and pushing geostatistical modeling even further into the arms of highperformance computing. It makes inferences for missing data by leveraging spacetime measurements of one or more fields. We propose a highperformance implementation of a widely applied spacetime model for largescale systems using a twolevel parallelization technique. At the inner level, we rely on stateoftheart dense linear algebra libraries and parallel runtime systems to perform complex matrix operations required to evaluate the maximum likelihood estimation (MLE). At the outer level, we parallelize the optimization process using a distributed implementation of the particle swarm optimization (PSO) algorithm. At this level, parallelization is accomplished using MPI subcommunicators, such that the nodes in each subcommunicator perform a single MLE iteration at a time. To evaluate the effectiveness of the proposed methodology, we assess the accuracy of the newly implemented spacetime model on a set of largescale synthetic spacetime datasets. Moreover, we use the proposed implementation to model two air pollution datasets from the Middle East and US regions with 550 spatial locations X730 time slots and 945 spatial locations X500 time slots, respectively. The evaluation shows that the proposed approach satisfies high prediction accuracy on both synthetic datasets and real particulate matter (PM) datasets in the context of the air pollution problem. We achieve up to 757.16 TFLOPS/s using 1024 nodes (75% of the peak performance) using 490K geospatial locations on ShaheenII Cray XC40 system.