Search NASA⌕ Search

SEARCH · Search NASA

Results for “Computationally efficient surrogate modeling”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 181 records · Page 10

Multi-fidelity information fusion with concatenated neural networks

Recently, computational modeling has shifted towards the use of statistical inference, deep learning, and other data-driven modeling frameworks. Although this shift in modeling holds promise in many applications like design optimization and real-time control by lowering the computational burden, training deep learning models needs a huge amount of data. This big data is not always available for scientific problems and leads to poorly generalizable data-driven models. This gap can be furnished by leveraging information from physics-based models. Exploiting prior knowledge about the problem at hand, this study puts forth a physics-guided machine learning (PGML) approach to build more tailored, effective, and efficient surrogate models. For our analysis, without losing its generalizability and modularity, we focus on the development of predictive models for laminar and turbulent boundary layer flows. In particular, we combine the self-similarity solution and power-law velocity profile (low-fidelity models) with the noisy data obtained either from experiments or computational fluid dynamics simulations (high-fidelity models) through a concatenated neural network. We illustrate how the knowledge from these simplified models results in reducing uncertainties associated with deep learning models applied to boundary layer flow prediction problems. The proposed multi-fidelity information fusion framework produces physically consistent models that attempt to achieve better generalization than data-driven models obtained purely based on data. While we demonstrate our framework for a problem relevant to fluid mechanics, its workflow and principles can be adopted for many scientific problems where empirical, analytical, or simplified models are prevalent. In line with grand demands in novel PGML principles, this work builds a bridge between extensive physics-based theories and data-driven modeling paradigms and paves the way for using hybrid physics and machine learning modeling approaches for next-generation digital twin technologies.

42 ENGINEERING↗

The confluence of machine learning and multiscale simulations

Multiscale modeling has a long history of use in structural biology, as computational biologists strive to overcome the time- and length-scale limits of atomistic molecular dynamics. Contemporary machine learning techniques, such as deep learning, have promoted advances in virtually every field of science and engineering and are revitalizing the traditional notions of multiscale modeling. Deep learning has found success in various approaches for distilling information from fine-scale models, such as building surrogate models and guiding the development of coarse-grained potentials. However, perhaps its most powerful use in multiscale modeling is in defining latent spaces that enable efficient exploration of conformational space. In conclusion, this confluence of machine learning and multiscale simulation with modern high-performance computing promises a new era of discovery and innovation in structural biology.

59 BASIC BIOLOGICAL SCIENCES↗

Calibration of cloud and aerosol related parameters for solar irradiance forecasts in WRF-solar

Model parameters are a major source of uncertainty in numerical weather prediction. Recently, the Weather Research and Forecasting model with Solar extensions (WRF-Solar) has been upgraded by enhancing the treatment of sub-grid scale cloud and aerosols with augmentations of a sub-grid scale cloud scheme (CLD3) and an upgraded aerosol-aware Thompson-Eidhammer scheme (TE14). However, the value of model parameters associated with these parameterizations are assigned based on limited measurements or theoretical calculations. Calibrating the most sensitive parameters has the potential to improve solar irradiance predictions. Here, we adopted a multiobjective surrogate-based optimization (SBO) framework to calibrate nine parameters used in CLD3 and TE14 that lead to the largest sensitivity in simulated irradiance. The normalized mean-absolute-error (NMAE) of global horizontal irradiance (GHI) and direct normal irradiance (DNI) are minimized by calibrating WRF-Solar over two regions including the Southern Great Plains (SGP) and Central California, in order to focus on parameter calibration under cloudy conditions with different aerosol loading. The results show that generalized linear model (GLM)-based surrogate models approximate physical models well, particularly when the third order and three-way interaction terms are considered. The SBO framework efficiently searches the parameter space for optimal solutions with less computational costs than directly calibrating the physical model. We first calibrate CLD3 parameters over the less-polluted SGP region. Optimized CLD3 parameters alone result in NMAE reduction by 14% for the site-mean and up to 33% for individual cases over the SGP region. With further calibration of TE14 parameters over the Central California during active fire periods, the optimized parameters lead to over 20% reductions of NMAE. Our investigation reveals, however, that optimizing TE14 has a limited impact on irradiance simulations under less-polluted conditions in the SGP.

14 SOLAR ENERGY↗

Accelerating the Computation of Detailed Chemical Reaction Kinetics for Simulating Combustion of Complex Fuels

Combustion of hydrocarbon fuels has been a very challenging scientific and engineering problem due to the complexity of turbulent flows and hydrocarbon reaction kinetics. There is an urgent need to develop an efficient modeling capability to accurately predict the combustion of complex fuels. Detailed chemical kinetic models for the surrogates of fuels such as gasoline, diesel and JP-8 consist of thousands of chemical species and Arrhenius reaction steps. Oxygenated fuels such as bio-fuels and heavier hydrocarbons, such as from newer fossil fuel sources, are expected to have a much more complex chemistry requiring increasingly larger chemical kinetic models. Such models are beyond current computational capability, except for homogeneous or partially stirred reactor type calculations. The advent of highly parallel multi-core processors and graphical processing units (GPUs) promises a steep increase in computational performance in the coming years. This paper will present a software framework that translates the detailed chemical kinetic models to high-performance code targeted for GPU accelerators.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Accelerating the Computation of Detailed Chemical Reaction Kinetics for Simulating Combustion of Complex Fuels

Combustion of hydrocarbon fuels has been a very challenging scientific and engineering problem due to the complexity of turbulent flows and hydrocarbon reaction kinetics. There is an urgent need to develop an efficient modeling capability to accurately predict the combustion of complex fuels. Detailed chemical kinetic models for the surrogates of fuels such as gasoline, diesel and JP-8 consist of thousands of chemical species and Arrhenius reaction steps. Oxygenated fuels such as bio-fuels and heavier hydrocarbons, such as from newer fossil fuel sources, are expected to have a much more complex chemistry requiring increasingly larger chemical kinetic models. Such models are beyond current computational capability, except for homogeneous or partially stirred reactor type calculations. The advent of highly parallel multi-core processors and graphical processing units (GPUs) promises a steep increase in computational performance in the coming years. This paper will present a software framework that translates the detailed chemical kinetic models to high- performance code targeted for GPU accelerators.

02 PETROLEUM↗

Transient uncertainty quantification and Global Sensitivity Analysis of the open-source Molten Chloride Reactor Experiment (MCRE) using GP-PCA surrogate models

Uncertainties in the thermophysical properties of molten salts impact both the steady-state and transient behavior of Molten Salt Reactors (MSRs). In this work, we aim to quantify the influence of such uncertainties on the transient operation of the Molten Chloride Reactor Experiment (MCRE), utilizing the open-source specifications provided for this reactor. Seven representative transient scenarios are considered. For each scenario, we evaluate the impact of thermophysical property uncertainties on four key multiphysics model output variables of interest (VoIs): maximum power density, maximum fuel temperature, maximum reflector temperature, and average fuel velocity magnitude. In addition, we perform a Global Sensitivity Analysis (GSA) by computing Sobol’ indices for the uncertain input parameters to determine their contribution to the variability of each VoI. Conducting GSA is computationally intensive due to the large number of required evaluations of the high-fidelity multiphysics model. To mitigate this cost, we develop a surrogate modeling framework that combines Gaussian Process (GP) regression with Principal Component Analysis (PCA), enabling efficient sample generation for the GSA. Our results show that for energy-related VoIs, thermal conductivity is the dominant contributor to uncertainty. In contrast, for flow-related VoIs, density and dynamic viscosity are the primary sources of uncertainty. The specific heat of the fuel salt was found to play a secondary role in the transient analyses.

42 - ENGINEERING↗

Circumventing data imbalance in magnetic ground state data for magnetic moment predictions

Abstract Magnetic materials play a crucial role in the transition to more sustainable forms of energy and electric vehicles. There is an anticipated shortage in magnetic materials in the future, and as a result there is an urgent need to discover and design new magnetic materials. Computational magnetic material design using density functional theory is daunting because of the challenge in identifying magnetic ground states from a combinatorially large set of possibilities. Machine learning offers a path forward by enabling efficient surrogate models that can more readily enumerate these states, but there is a dearth of training data available, and what is available tends to be imbalanced with too much non-magnetic data. In this work we show that the discrete and previously tackled data imbalance that exists at the level of the magnetic ordering leads to an imbalanced continuous distribution with many zeros when the data is unraveled at the atomic magnetic moment level, which subsequently leads to models with low accuracy for magnetic properties. We mitigate this by using a two-part model framework. Our scheme is able to classify atoms into magnetic and non-magnetic with an F1 score and Matthew’s correlation coefficient (MCC) of ~91% and then to provide an implicit embedding representation that maps directly onto the magnitude of the magnetic moment with a mean absolute error of 0.1 μ B . Beyond screening for new magnetic materials, we demonstrate an additional practical use case of our scheme: the provision of good initial guesses for magnetic moments in first-principles electronic relaxations. Such initialization is shown to lead to faster convergence to configurations that lie closer to the ground state.

Computer Science↗

Interpolant Improvements and Lessons Learned

This presentation is for the OpenMDAO workshop 2022 and updates users on recent improvements to interpolant methods. Specifically, we discuss computational improvements, visualization capabilities, and suggested best practices for using interpolants. The term interpolants is used synonymously with metamodels and surrogate models. The goal of the presentation is to increase adoption of efficient interpolant methods and increase users’ awareness to built-in features within OpenMDAO.

Multidisciplinary Design Optimization↗

Deep Learning-Assisted Real-Time Forward Modeling of Electromagnetic Logging in Complex Formations

Higher dimensional (i.e., 2-D and 3-D) modeling is indispensable to correctly evaluate the responses of electromagnetic logging tools in complex formation environments. However, limited by the high computational cost of rigorous modeling such as finite difference method and the finite element method, the real-time applications in the well logging industry primarily rely on the 1-D forward solver, which would result in erroneous formation evaluation for complex scenarios. As a result, aiming at realizing fast modeling for electromagnetic logging tools in complex formations, this paper proposes a general framework assisted by deep neural networks (DNNs). The framework consists of three modules: earth model classification, parameter extraction, and surrogate construction. Separate DNNs are trained and tested for different modules. The accuracy and efficiency of the DNN assisted fast modeling are validated by several experiments. Here, this study finds that the fast modeling assisted by DNNs is able to calculate the tool responses and reconstruct the subsurface formations in real-time.

97 MATHEMATICS AND COMPUTING↗

Model-based, in-situ, non-destructive qualification and certification of parts made by autonomous additive manufacturing

To address the significant productivity challenges associated with the qualification and certification (Q&C) tasks of additively manufactured (AM) parts, which have traditionally relied on rigorous post‐build inspection and testing, we propose an integrated framework that combines model‐based qualification and certification (MBQ&C) with autonomous additive manufacturing (AAM). MBQ&C employs high‐fidelity predictive models, developed within the Integrated Computational Materials Engineering (ICME) paradigm, to simulate process–structure–property–performance relationships for assessing a part’s fitness for use. Since predictive models are commonly machine learning (ML)-based or reduced-order surrogates of validated physics models, they run efficiently, enabling timely inference. In parallel, the self-driving AAM utilises ML-based adaptive, closed‐loop control strategies to avoid, mitigate, or repair defects and anomalies during fabrication, thereby increasing the likelihood of producing acceptable parts. A key feature of the combined AAM-MBQ&C framework is that predictive models explicitly incorporate defects or anomalies that persist after the build, using instance-specific data captured via in-situ sensing. This customisation enables a build‐specific assessment of fitness for use, rather than relying on nominal or generic parameters. Such individualised evaluation provides a robust basis for Q&C-related acceptance decisions relating to each build. Additionally, the rapid solution capabilities of ML or reduced-order models enable the determination of a part’s suitability for service shortly after build completion. As the framework matures, it has the potential to substantially reduce reliance on conventional point‐design approaches—such as time‐consuming post‐build computed tomography scanning and costly destructive testing. Thus, the AAM-MBQ&C framework represents a transformative, scalable strategy for quality assurance of AM components, as parts produced within a stable, validated, and certified envelope can be certified with reduced testing. Key benefits include: (1) significant gains in Q&C productivity through efficient, model-centric assessment; (2) performance-based classification of defects into critical and non-critical categories; (3) the ability to predict potential deviations in the performance of parts affected by real-time, adaptive process control interventions relative to those produced under a certified process, and (4) the enabling of virtual Q&C for service environments that are difficult, hazardous, or impractical to access or reproduce experimentally. Collectively, these capabilities strengthen the business case for AM, particularly for high‐consequence and mission‐critical applications. Finally, although this work focuses on powder-based AM, the proposed techniques could be extended to AM processes employing alternative feedstock forms.

Gunasegaram, Dayalan↗

An adaptive importance sampling algorithm for Bayesian inversion with multimodal distributions

Parametric uncertainties are encountered in the simulations of many physical systems, and may be reduced by an inverse modeling procedure that calibrates the simulation results to observations on the real system being simulated. Following Bayes' rule, a general approach for inverse modeling problems is to sample from the posterior distribution of the uncertain model parameters given the observations. However, the large number of repetitive forward simulations required in the sampling process could pose a prohibitive computational burden. This difficulty is particularly challenging when the posterior is multimodal. We present in this paper an adaptive importance sampling algorithm to tackle these challenges. Two essential ingredients of the algorithm are: 1) a Gaussian mixture (GM) model adaptively constructed as the proposal distribution to approximate the possibly multimodal target posterior, and 2) a mixture of polynomial chaos (PC) expansions, built according to the GM proposal, as a surrogate model to alleviate the computational burden caused by computational-demanding forward model evaluations. In three illustrative examples, the proposed adaptive importance sampling algorithm demonstrates its capabilities of automatically finding a GM proposal with an appropriate number of modes for the specific problem under study, and obtaining a sample accurately and efficiently representing the posterior with limited number of forward simulations.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

A trust-region algorithm for the optimization of PSA processes using reduced-order modeling

The last few decades have seen a considerable increase in the applications of adsorptive gas separation technologies, such as pressure swing adsorption (PSA); the applications range from bulk separations to trace contaminant removal. PSA processes are based on solid-gas equilibrium and operate under periodic transient conditions [1]. Bed models for these processes are therefore defined by coupled nonlinear partial differential and algebraic equations (PDAEs) distributed in space and time with periodic boundary conditions that connect the processing steps together and high nonlinearities arising from non-isothermal effects and nonlinear adsorption isotherms. As a result, the optimization of such systems for either design or operation represents a significant computational challenge to current nonlinear programming algorithms. Model reduction is a powerful methodology that permits systematic generation of cost-efficient low-order representations of large-scale systems that result from discretization of such PDAEs. In particular, low-dimensional approximations can be obtained from reduced order modeling (ROM) techniques based on proper orthogonal decomposition (POD) and can be used as surrogate models in the optimization problems. In this approach, a representative ensemble of solutions of the dynamic PDAE system is constructed by solving a higher-order discretization of the model using the method of lines, followed by the application of Karhunen-Loeve expansion to derive a small set of empirical eigenfunctions (POD modes). These modes are used as basis functions within a Galerkin's projection framework to derive a low-order DAE system that accurately describes the dominant dynamics of the PDAE system. This approach leads to a DAE system of significantly lower order, thus replacing the one obtained from spatial discretization before and making optimization problem computationally efficient [2]. The ROM methodology has been successfully applied to a 2-bed 4-step PSA process used for separating a hydrogen-methane mixture in [3]. The reduced order model developed was successfully used to optimize this process to maximize hydrogen recovery within a trust-region. We extend this approach in this work to develop a rigorous trust-region algorithm for ROM-based optimization of PSA processes. The trust-region update rules and sufficient decrease condition for the objective is used to determine the size of the trust-region. Based on the decrease in the objective function and error in the ROM, a ROM updation strategy is designed [4, 5]. The inequalities and bounds are handled in the algorithm using exact penalty formulation, and a non-smooth trust-region algorithm by Conn et al. [6] is used to handle non-differentiability. To ensure that the first order consistency condition is met and the optimum obtained from ROM-based optimization corresponds to the optimum of the original problem, a scaling function, such as one proposed by Alexandrov et al. [7], is incorporated in the objective function. Such error control mechanism is also capable of handling numerical inconsistencies such as unphysical oscillations in the state variable profiles. The proposed methodology is applied to optimize a PSA process to concentrate CO{sub 2} from a nitrogen-carbon dioxide mixture. As in [3], separate ROMs are developed for each operating step with different POD modes for each state variable. Numerical results will be presented for optimization case studies which involve maximizing CO{sub 2} recovery, feed throughput or minimizing overall power consumption.

97 MATHEMATICS AND COMPUTING↗

A general framework for quantifying uncertainty at scale

Abstract In many fields of science, comprehensive and realistic computational models are available nowadays. Often, the respective numerical calculations call for the use of powerful supercomputers, and therefore only a limited number of cases can be investigated explicitly. This prevents straightforward approaches to important tasks like uncertainty quantification and sensitivity analysis. This challenge can be overcome via our recently developed sensitivity-driven dimension-adaptive sparse grid interpolation strategy. The method exploits, via adaptivity, the structure of the underlying model (such as lower intrinsic dimensionality and anisotropic coupling of the uncertain inputs) to enable efficient and accurate uncertainty quantification and sensitivity analysis at scale. Here, we demonstrate the efficiency of this adaptive approach in the context of fusion research, in a realistic, computationally expensive scenario of turbulent transport in a magnetic confinement tokamak device with eight uncertain parameters, reducing the effort by at least two orders of magnitude. In addition, we show that this refinement method intrinsically provides an accurate surrogate model that is nine orders of magnitude cheaper than the high-fidelity model.

Farcaş, Ionuţ-Gabriel (ORCID:0000000238841580)↗

NUMERICAL INVESTIGATION OF FUEL PROPERTY EFFECTS ON MIXED-MODE COMBUSTION IN A SPARK-IGNITION ENGINE

In the present study, mixed-mode combustion of an E30 fuel in a direct-injection spark-ignition engine is numerically investigated at a fuel-lean operating condition using multidimensional computational fluid dynamics (CFD). A fuel surrogate matching Research Octane Number (RON) and Motor Octane Number (MON) of E30 is first developed using neural network based non-linear regression model. To enable efficient 3D engine simulations, a 164-species skeletal reaction mechanism incorporating NO, chemistry is reducedfrom a detailed chemical kinetic model. A hybrid approach that incorporates the G-equation model for tracking turbulentflame front, and the multi-zone well-stirred reactor model for predicting auto-ignition in the end gas, is employed to account for turbulent combustion interactions in the engine cylinder. Predicted in-cylinder pressure and heat release rate traces agree well with experimental measurements. The proposed modelling approach also captures moderated cyclic variability. Two different types of combustion cycles, corresponding to purely deflagrative and mixed-mode combustion, are observed. In contrast to the purely deflagrative cycles, mixed-mode combustion cycles feature early flame propagation followed by end-gas auto-ignition, leading to two distinctive peaks in heat release rate traces. The positive correlation between mixed-mode combustion cycles and early flame propagation is well captured by simulations. With the validated numerical setup, effects of NOx chemistry on mixed-mode combustion predictions are investigated. NOx chemistry is found to promote auto-ignition through residual gas recirculation, while the deflagrative flame propagation phase remains largely unaffected. Local sensitivity analysis is then performed to understand effects of physical and chemical properties of the fuel, i.e., heat of evaporation (HoV) and laminar flame speed (S-L). An increased HoV tends to suppress end-gas auto-ignition due to increased vaporization cooling, while the impact of HoV on flame propagation is insignificant. In contrast, an increased SL is found to significantly promote both flame propagation and auto-ignition. The promoting effect of SL on auto-ignition is not a direct chemical effect; it is rather caused by an advancement of the combustion phasing, which increases compression heating of the end gas.

detailed chemical kinetics↗

Using a surrogate-assisted Bayesian framework to calibrate the runoff-generation scheme in the Energy Exascale Earth System Model (E3SM) v1

Abstract. Runoff is a critical component of the terrestrial water cycle, and Earth system models (ESMs) are essential tools to study its spatiotemporal variability. Runoff schemes in ESMs typically include many parameters so that model calibration is necessary to improve the accuracy of simulated runoff. However, runoff calibration at a global scale is challenging because of the high computational cost and the lack of reliable observational datasets. In this study, we calibrated 11 runoff relevant parameters in the Energy Exascale Earth System Model (E3SM) Land Model (ELM) using a surrogate-assisted Bayesian framework. First, the polynomial chaos expansion machinery with Bayesian compressed sensing is used to construct computationally inexpensive surrogate models for ELM-simulated runoff at 0.5∘ × 0.5∘ for 1991–2010. The error metric between the ELM simulations and the benchmark data is selected to construct the surrogates, which facilitates efficient calibration and avoids the more conventional, but challenging, construction of high-dimensional surrogates for the ELM simulated runoff. Second, the Sobol' index sensitivity analysis is performed using the surrogate models to identify the most sensitive parameters, and our results show that, in most regions, ELM-simulated runoff is strongly sensitive to 3 of the 11 uncertain parameters. Third, a Bayesian method is used to infer the optimal values of the most sensitive parameters using an observation-based global runoff dataset as the benchmark. Our results show that model performance is significantly improved with the inferred parameter values. Although the parametric uncertainty of simulated runoff is reduced after the parameter inference, it remains comparable to the multimodel ensemble uncertainty represented by the global hydrological models in ISMIP2a. Additionally, the annual global runoff trend during the simulation period is not well constrained by the inferred parameter values, suggesting the importance of including parametric uncertainty in future runoff projections.

58 GEOSCIENCES↗

An Efficient Bayesian Method for Advancing the Application of Deep Learning in Earth Science

Despite their wide and successful applications, deep learning (DL) models are prone to overfitting for small training datasets, produce a poor predictive performance for uncertain data, and provide point estimations without any indication of the accuracy and credibility. These limitations of the deterministic DL models hinder their effective application in Earth system science where the labelled data are sparse, noisy and incomplete with large uncertainty and where the predictive uncertainty quantification is needed for scientific understanding and policy decision making. Integration of Bayesian inference into DL models adds an estimate of uncertainty and regularization in the predictions. However, traditional Bayesian methods are computationally unaffordable and inflexible for high-dimensional problems, which limits their application to DL systems that typically have millions of model parameters. In this effort, we propose an efficient and general-purpose Bayesian inference method to advance DL model optimization and uncertainty quantification, so as to facilitate the adoption of DL in Earth sciences. In a demonstration, we integrate the proposed Bayesian method with a feedforward neural network (NN) to build a fast-to-evaluate surrogate of the complex Energy Exascale Earth System Land Model for efficient modeling. The formulated Bayesian NN, using a small number of training data, produces an accurate prediction with high credibility, whereas with the same small training size, the deterministic NN cannot yield a reasonable estimation and does not provide confidence information. The proposed Bayesian method is computationally efficient and flexible, capable of integration with diverse network variants such as convolutional NNs and recurrent NNs to advance the application of DL in Earth sciences.

Lu, Dan↗

Real-time capable modeling of ICRF heating on NSTX and WEST via machine learning approaches

Abstract A real-time capable core Ion Cyclotron Range of Frequencies (ICRF) heating model on NSTX and WEST is developed. The model is based on two nonlinear regression algorithms, the random forest ensemble of decision trees and the multilayer perceptron neural network. The algorithms are trained on TORIC ICRF spectrum solver simulations of the expected flat-top operation scenarios in NSTX and WEST assuming Maxwellian plasmas. The surrogate models are shown to successfully capture the multi-species core ICRF power absorption predicted by the original model for the high harmonic fast wave and the ion cyclotron minority heating schemes while reducing the computational time by six orders of magnitude. Although these models can be expanded, the achieved regression scoring, computational efficiency and increased model robustness suggest these strategies can be implemented into integrated modeling frameworks for real-time control applications.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

A Physics-Constrained Deep Learning Model for Simulating Multiphase Flow in 3D Heterogeneous Porous Media

Physics-based simulators for multiphase flow in porous media emulate nonlinear processes with coupled physics, and usually require extensive computational resources for software development, maintenance and simulation execution. As a result, a huge demand exists for fast modeling of coupled processes in a wide range of subsurface applications including geological sequestration, hydrocarbon recovery and geothermal energy extraction. In this work, an efficient physics-constrained deep learning model is developed for solving multiphase flow in 3-Dimensional (3D) heterogeneous porous media. The model fully leverages the spatial topology predictive capability of convolutional neural networks, specifically U-Net with successive contracting and expansive steps, and is coupled with an efficient continuity-based smoother to predict flow responses that need spatial continuity. Furthermore, the transient regions are penalized to steer the training process such that the model can accurately capture flow in these regions. The model takes inputs including properties of porous media, fluid properties and well controls, and predicts the temporal-spatial evolution of the state variables (pressure and saturation). While maintaining the continuity of fluid flow, the 3D spatial domain is decomposed into 2D images for reducing training cost, and the decomposition results in an increased number of training data samples and better training efficiency. Additionally, a surrogate model is separately constructed as a postprocessor to calculate well flow rate based on the predictions of state variables from the deep learning model. We use the example of CO 2 injection into saline aquifers, and apply the physics-constrained deep learning model that is trained from physics-based simulation data and emulates the physics process. The model performs prediction with a speedup of ~ 1400 times compared to physics-based simulations, and the average temporal errors of predicted pressure and saturation plumes are 0.27% and 0.099% respectively. Furthermore, water production rate is efficiently predicted by a surrogate model for well flow rate, with a mean error less than 5%. Therefore, with its unique scheme to cope with the fidelity in fluid flow in porous media, the physics-constrained deep learning model can become an efficient predictive model for computationally demanding inverse problems or other coupled processes.

58 GEOSCIENCES↗