FAFaruk Omer Alpak

Faruk Omer Alpak

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

Faruk Omer Alpak has a total of 45 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Quantifying the Uncertainty of the 3D Evolution of CO2-Mass and Pressure Plumes for Site Screening Applications with Physics-Informed Fourier Neural Operators

    by Suraj Pawar, Anirban Chandra, Aniruddha Panda, Faruk O. Alpak, Jeroen Snippe, Pandu Devarakota, Marius Koch, Detlef Hohl

    Summary: Identification of the most suitable set of target storage sites in terms of capacity and containment early in the project phase is one of the key elements for successful carbon dioxide (CO2) sequestration. There is a need for rapid approximate models that can forecast the migration of CO2-mass and pressure plumes. One family of surrogate models that has proved effective for simulating CO2 plume migration is Fourier neural operator (FNO)-based techniques. We discuss the development of two FNO architectures, standard FNO (3DFNO) and U-Net augmented FNO (U-FNO), to serve as fast simulation engines in a CO2 storage site screening platform using deep neural networks and operator learning. We also discuss various physics embedding training strategies (e.g., mass conservation) for improving the reliability of these artificial intelligence simulations. We provide computational results in 3D plume-migration evolution problems in terms of training costs and prediction accuracy for the 3DFNO and U-FNO protocols. We focus our investigations on layered geology problems typical of the screening stage. In inference mode, both models offer O(105) computational acceleration with respect to conventional reservoir simulation subject to an acceptable level of compromise in accuracy. While U-FNO incurs approximately 50% higher computational costs compared with 3DFNO, it improves the prediction accuracy. We demonstrate that U-FNO yields more accurate predictions than 3DFNO in layered systems for CO2-mass accumulation and CO2-rich phase (gas) saturation, while 3DFNO yields more accurate predictions for evolution of pressure change. The platform over which the FNO engines are deployed is designed for predicting the 3D spatiotemporal evolution of CO2-mass accumulation, gas phase saturation, and pressure in arbitrarily layered systems. The fast simulation platform enables Monte Carlo sampling to compute statistical distribution functions of plume-migration performance indicators. Fast FNO-based surrogate models thus enable a previously impossible Monte Carlo–based uncertainty quantification workflow for predicting CO2-mass plume-migration distance and post-injection pressure evolution.


  • Application of Embed-to-Control Observe Deep-Learning-Based Reservoir Surrogate in CO2 Storage Monitoring and Multiobjective Optimization

    by Yuhao Chen, Quang Minh Nguyen, Usman Abdulkareem, Mustafa Onur, Faruk Omer Alpak

    Summary: The focus of this paper is on the use of a data-driven method for predicting reservoir performance in a carbon dioxide (CO2) storage application. The aim is to forecast state variables and well outputs for a compositional fluid system, as well as to perform multiobjective optimization under geological uncertainty, all by using a deep learning (DL)-based surrogate model. The surrogate model used in this work is based on the multimodel embed-to-control observe (E2CO) architecture, consisting of four blocks of neural networks (NNs), which are encoder, transition, transition output, and decoder. Our previous studies published by Nguyen et al. 2023bNguyen et al. 2023c,47 considered bi-objective optimization under geological uncertainty of CO2 storage, but using a high-fidelity simulator (HFS). Also, the previous multimodel E2CO proxy models published in our studies by Nguyen and Onur 2024aNguyen and Onur 2024b and 49 and Nguyen et al. (2024) are valid for single-objective reservoir production optimization using the E2CO proxy built on the training samples from two-phase black oil reservoir models. This is the first study for implementing E2CO, trained by samples from a compositional model based on SPE10 for a bi-objective optimization problem of CO2 storage. In this study, we combine the surrogate model with the stochastic-gradient-based line-search sequential quadratic programming (LS-SQP) framework and implement the lexicographic method to handle robust multiobjective optimization with the nonlinear state constraints. Five realizations of the SPE10 sector model, consisting of different channelized heterogeneous distributions with eight vertical wells, are used to demonstrate the prediction and optimization performances of the proposed framework. The framework fully leverages various results and insights from previous studies, allowing us to achieve excellent optimization (with nonlinear state constraint) results at very affordable computational expenses, while still maintaining good agreement with the results from the HFS.


  • Embed-to-Control-Based Deep Learning Surrogate for Robust Nonlinearly Constrained Life Cycle Production Optimization: A Realistic Deepwater Application

    by Quang Minh Nguyen, Mustafa Onur, Faruk Omer Alpak

    Summary: This paper presents a realistic deepwater application of a deep learning (DL)-based reservoir surrogate model in forecasting reservoir performance (via the prediction of state variables and well outputs) and nonlinearly constrained production optimization under geological uncertainty by utilizing an efficient gradient-based method. The reservoir surrogate model is based on the multimodel embed-to-control observe (E2CO) architecture, consisting of four blocks of neural networks (NNs): encoder, transition, transition output, and decoder. In this work, the surrogate model is coupled with the powerful stochastic-gradient-based line-search sequential quadratic programming (LS-SQP) workflow to handle robust production optimization in the presence of nonlinear state constraints. This is the first time the E2CO-based production optimization method driven by a novel nonlinear optimization is applied to a reservoir model with very realistic characteristics of a deepwater reservoir (not a toy reservoir model) in terms of structural architecture and flow property distribution in an efficient data-driven manner using the multimodel E2CO DL surrogate as detailed by Nguyen and Onur (2025). Five realizations of the synthetic realistic Reservoir X model with different channelized heterogeneous permeability distributions under waterflooding are used for showcasing the prediction and optimization performances of the DL-based framework. The results from this framework are directly and quantitatively compared with those simulated using a commercial high-fidelity reservoir simulator (HFS).

  • Nonlinearly Constrained Life-Cycle Production Optimization Under Geological Uncertainty: A Realistic Application

    by Omer Lutfu Toktas, Quang Minh Nguyen, Mustafa Onur, Faruk Omer Alpak

    Summary: This study presents a realistic application of our previously developed nonlinearly constrained stochastic gradient-based optimization framework that can efficiently solve nonlinearly constrained robust production optimization problems. We apply the optimization framework to a synthetic, yet realistic, field optimization problem using multiple history-matched realizations of the geologic model. For the synthetic realistic field example, the reservoir is assumed to undergo production for several years. Well controls (bottomhole pressures, or BHPs, for producers and water injection rates for injectors) on a sequence of control steps (time intervals) are optimized during the forecast period by maximizing the expected value of the net present value (NPV, or ensemble average) over multiple realizations of subsurface description for the subsequent 10 years of production under a waterflooding scenario subject to nonlinear constraints. The optimization framework is based on our in-house nonlinear optimizer employing sequential quadratic programming (SQP) coupled with stochastic simplex approximated gradients (StoSAG). It incorporates the line-search procedure within SQP and is referred to as line-search sequential quadratic programming (LS-SQP). We couple LS-SQP with two different constraint handling schemes: the expected value constraint scheme and the minimum-maximum (min-max) constraint scheme. Both bound constraints on well controls and nonlinear constraints on field liquid production rate (FLPR), field water injection rate FWIR, and/or individual producer water production rates are considered. We consider several different history-matched geological realizations of the realistic field example with several injectors and producers. A black-oil commercial reservoir simulator is utilized in the optimization workflow to assess the objective function with imposed constraints. Results show that the framework LS-SQP algorithm with StoSAG can effectively handle the nonlinear constraints in a robust life cycle production optimization problem. Moreover, our results show that the LS-SQP framework with any of the two different constraint-handling schemes considered effectively handles the nonlinear constraints in a life cycle robust production optimization problem. However, the expected value constraint scheme results in a higher optimal NPV than the min-max constraint scheme, but at the cost of possible constraint violation for some individual geological realizations. Application results demonstrate that our nonlinearly constrained stochastic gradient-based robust optimization framework has great potential to be a rapid approximate capability for performing life cycle production optimization under subsurface uncertainty with bound and nonlinear constraints.


  • Physics-Based Reduced-Order Model Preconditioning for Carbon Capture and Sequestration Simulations

    by Faruk Omer Alpak, Jeroen Snippe, Antonio Crapsi

    Summary: A novel on-the-fly version of the physics-based reduced-order model (ROM) preconditioning technique is developed for carbon capture and sequestration (CCS) simulations. The preconditioner leverages the proper orthogonal decomposition (POD) method to construct a ROM approximation of the linearized pressure equation. Using a small set of precomputed pressure solutions learned during the flow simulation, the pressure block is projected into a lower-order system in the POD-ROM preconditioner. On-the-fly training and inference during a single simulation eliminates the need for prior training. We propose strategies for the optimal selection of training and inference periods by exploiting prior knowledge about CCS flow physics (e.g., relatively shorter injection and longer plume-migration periods). The POD-ROM preconditioner is validated and evaluated against popular alternatives such as the constrained pressure residual algebraic multigrid (CPR-AMG) and block incomplete linear unit (BILU) preconditioners in realistic but synthetic CCS simulation models. Various CCS-relevant physical phenomena (e.g., thermal effects, dissolution, evaporation, and diffusion) and training and inference period schedules have been investigated for the POD-ROM preconditioner’s on-the-fly inference paradigm. We fine-tune the balance between the number of POD basis functions and the necessary linear iterations in the investigations. We also provide recommendations for the training and inference intervals and their optimal selection. Findings from representative CCS simulations indicate that utilizing the on-the-fly POD-ROM preconditioner accelerates the convergence of the iterative linear solution step compared with the traditional CPR-AMG and BILU preconditioning methods. Time spent in the linear solver improves accordingly, exceeding a speedup factor of two with the POD-ROM preconditioner over CPR-AMG. This, in turn, gives rise to overall simulation time speedups in the range of 1.1–1.8 times “out-of-the-box” in reference to CPR-AMG without any further tuning. We show that the POD-ROM preconditioner yields cost-effective linear solves while still providing accurate estimates of the original high-order system’s solution.

  • Deep Learning–Driven Acceleration of Stochastic Gradient Methods for Well Location Optimization under Uncertainty

    by Esmail Eltahan, Kamy Sepehrnoori, Faruk Omer Alpak

    Summary: We have developed the deep learning-accelerated gradient (DLAG) algorithm, a novel solution for well location optimization (WLO) problems. DLAG accelerates optimization by leveraging data from the explored parameter space to train a deep neural network (NN) that maps well locations to the objective function. The algorithm utilizes the analytical gradient of the NN, resulting in an effective search direction while saving the computational cost associated with stochastic gradient perturbations. Our method introduces a novel NN architecture, the spatial pairwise interaction network (SPINet), composed of independent neural pathways (NPs) to capture primary well characteristics and contextual NPs to capture complex interactions with the neighboring wells. For the contextual NP, we explore using the popular attention mechanism and simpler mechanism with weight-sharing multilayer perceptron (MLP) layers. To evaluate the architectures, we designed the Bird Ensemble (BE) test problem, which resembles the structure of the WLO problem. SPINet structure proves to efficiently model pairwise relationships while maintaining invariance to input permutation. The NP with a weight-sharing mechanism has superior performance in terms of mean square error and gradient accuracy while also demonstrating more consistent and reliable training outcomes. For the test problem, DLAG dramatically improved computational efficiency, reducing function evaluations by one to two orders of magnitude. Following this validation, we successfully applied our method to optimize trajectories of injection and production wells in a modified version of the Egg reservoir model. To alleviate random artifacts and the inherent possibility of encountering local minima, we executed 20 distinct instances of the WLO problem, comparing performance with and without DLAG. On average, our DLAG optimization method proved significantly more efficient than the traditional approach. This efficiency gain is particularly impactful for real-world applications where project timelines are tightly constrained.


  • A Quasi-Newton Trust-Region Method for Well Location Optimization Under Uncertainty

    by Esmail Eltahan, Faruk Omer Alpak, Kamy Sepehrnoori

    Summary: Subsurface development involves well-placement decisions considering the highly uncertain understanding of the reservoir in the subsurface. The simultaneous optimization of a large number of well locations is a challenging problem. Traditional gradient-based methods can be adapted for well location optimization (WLO) when these problems are converted into real-valued representations and equipped with protocols to handle noisy objective functions. However, their application to large-scale scenarios often remains impractical. This impracticality arises because computing gradients of the objective function can be prohibitively expensive in realistic settings, particularly without using the adjoint method. In this paper, we explore the application of a novel quasi-Newton trust-region (TR) method that employs the stochastic simplex approximate gradient (StoSAG).We have implemented the Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton updating algorithm together with line-search (LS) and TR optimization strategies. The StoSAG-based optimization methods have been applied to a realistic synthetic reservoir featuring 26 wells considering two distinct cases: Each includes 20 realizations of porosity and permeability. The first case exhibits mild heterogeneity, while the second exhibits significant heterogeneity with a large correlation length. We have conducted a series of runs to evaluate the performance of these algorithms in addition to comparisons to the finite-difference (FD) and particle-swarm-optimization (PSO) algorithms. We introduce a novel approach to enhance the accuracy of StoSAG gradients by proposing modified StoSAG formulations. These formulations are tailored to exploit the structure of the objective function and to capture the relationships between its components and the individual optimization parameters. This approach involves using a correction matrix W informed by problem-specific knowledge. The entries of W vary from 0 to 1 and are proportional to the interference effects the neighboring wells have on each other. We determine those entries (or weights) based on the radii of investigation around the wells and the distance between the well pairs.Results indicate that the steepest-descent (SD) algorithm coupled with StoSAG has superior performance to PSO and FD. Although the objective function is prone to numerical noise and not continuously differentiable with respect to well locations, StoSAG overcomes this challenge because it acts as a smooth approximation. Comparative tests further confirm that the TR-BFGS is more effective than the LS-BFGS. Moreover, we show that using the proposed gradient correction procedure results in a significant acceleration in convergence, indicating an enhancement in the StoSAG gradient approximation quality. This enhancement allows the TR-BFGS algorithm to achieve considerably higher performance than SD, illustrating that the accuracy of the BFGS approximation benefits from improved gradient quality.


  • Bi-Objective Optimization of Subsurface CO2 Storage with Nonlinear Constraints Using Sequential Quadratic Programming with Stochastic Gradients

    by Quang Minh Nguyen, Mustafa Onur, Faruk Omer Alpak

    Summary: This study focuses on carbon capture, utilization, and sequestration (CCUS) via the means of nonlinearly constrained production optimization workflow for a CO2-enhanced oil recovery (EOR) process, in which both the net present value (NPV) and the net present carbon tax credits (NPCTC) are bi-objectively maximized, with the emphasis on the consideration of injection bottomhole pressure (IBHP) constraints on the injectors, in addition to field liquid production rate (FLPR) and field water production rate (FWPR), to ensure the integrity of the formation and to prevent any potential damage during the life cycle injection/production process. The main optimization framework used in this work is a lexicographic method based on the line-search sequential quadratic programming (LS-SQP) coupled with stochastic simplex approximate gradients (StoSAG). We demonstrate the performance of the optimization algorithm and results in a field-scale realistic problem, simulated using a commercial compositional reservoir simulator. Results show that the workflow can solve the single-objective and bi-objective optimization problems computationally efficiently and effectively, especially in handling and honoring nonlinear state constraints imposed onto the problem. Various numerical settings have been experimented with to estimate the Pareto front for the bi-objective optimization problem, showing the trade-off between the two objectives of NPV and NPCTC. We also perform a single-objective optimization on the total life cycle cash flow, which is the aggregated quantity of NPV and NPCTC, and quantify the results to further emphasize the necessity of performing bi-objective production optimization, especially when used in conjunction with commercial flow simulators that lack the capability of computing adjoint-based gradients.


  • Augmenting Deep Residual Surrogates with Fourier Neural Operators for Rapid Two-Phase Flow and Transport Simulations

    by Faruk O. Alpak, Janaki Vamaraju, James W. Jennings, Suraj Pawar, Pandu Devarakota, Detlef Hohl

    Summary: Accurate numerical modeling of multiphase flow and transport mechanisms is essential to study varied, complex physical phenomena including flow in subsurface oil and gas reservoirs and subsurface aquifers subject to CO2 sequestration. State-of-the-art complete physics-based solvers suffer from many computational challenges. High-fidelity data-driven surrogate models that solve the governing partial differential equations (PDEs) have the potential to optimize the time to solution and increase confidence in critical business and engineering decisions through better quantification of solution statistics. We leverage the recently proposed Fourier neural operators (FNOs) with quasilinear time complexity to capture the spectral information from feature maps to solve the coupled porous flow and transport PDEs. Embedding Fourier layers within the residual blocks results in a highly effective structure that, while achieving competitive accuracy, also enables efficient training of deeper networks with a dramatically reduced number of trainable parameters. The resulting novel deep-learning (DL) architecture is coined as FResNet++. FResNet++ uses squeeze and excitation blocks, atrous spatial pyramid pooling (ASPP), and attention blocks to increase its sensitivity to the relevant features and capture multiscale information, and it is specifically tuned to operate optimally to learn from and predict numerically simulated flow (pressure and saturation) fields. We demonstrate the ability of FResNet++ to generalize over multiple high-dimensional input parameter spaces that describe subsurface permeability and porosity heterogeneity. The resulting DL architecture accurately captures the complex interplay between viscous forces and highly heterogeneous permeability and porosity fields. We investigate two-phase flow in porous media, which is the archetypal problem for reservoir simulation giving rise to a system of nonlinearly coupled PDEs with highly heterogeneous coefficients. We show in blind tests that FResNet++ predicts saturation fields more accurately compared to ResU-Net and original FNO with fully connected linear layers. We additionally investigate the effects of using alternative loss functions and an alternative way of utilizing FResNet++ to increase its effectiveness. For the first time in the literature, we show that the spatiotemporal evolution of pressure and saturation fields can be jointly predicted with good accuracy using a single FResNet++ network over long time horizons in response to previously unseen permeability and porosity fields. After a moderate training investment on graphics processing units (GPUs), FResNet++ yields a speedup of at least four orders of magnitude compared to a conventional numerical PDE solver and operates with notably fewer trainable parameters compared to the original FNO. Our numerical experiments validate that FNOs can be utilized in various convolutional neural network-based architectures and can effectively substitute for repetitive physics-based forward simulations for scenario testing.


  • Open Access

    Correction to: Estimating permeability of 3D micro-CT images by physics-informed CNNs based on DNS

    by Stephan Gärttner, Faruk O. Alpak, Andreas Meier, Nadja Ray, Florian Frank