東華大學圖書館 |

Information-Theoretic Approach for Upscaling.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Information-Theoretic Approach for Upscaling./
作者:	Belivanis, Dimitrios Ioannis.
面頁冊數:	1 online resource (108 pages)
附註:	Source: Dissertations Abstracts International, Volume: 84-05, Section: A.
Contained By:	Dissertations Abstracts International84-05A.
標題:	Software. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29756186click for full text (PQDT)
ISBN:	9798357512581

Information-Theoretic Approach for Upscaling.
Belivanis, Dimitrios Ioannis.

Information-Theoretic Approach for Upscaling. - 1 online resource (108 pages)

Source: Dissertations Abstracts International, Volume: 84-05, Section: A.

Thesis (Ph.D.)--Stanford University, 2022.

Includes bibliographical references

Recent advances in hardware, algorithms, and scientific computing open new possibilities to revisit long-standing problems with additional tools in our arsenal. Examples of such problems, which are explored in this study, are inverse problems and upscaling of dynamic models. Both problems include nonlinear and functional minimization of the discrepancy between data and model predictions. This study develops an information-theoretic approach for upscaling of dynamic models affected by uncertainty. An optimal, time-dependent, probabilistic characterization of the macroscale model is obtained to yield minimum discrepancy with respect to assigned quantities of interest provided by the microscale solution. Hard data forming a training set are obtained at the fine scale from repeated synthetic simulations or from observations. Fine-scale information is transferred to the coarse scale via minimization of a loss function that consists of the cumulative average discrepancy, regularized by the cumulative exchange of information measured via mutual information. This general procedure is applied to transient flow processes in heterogeneous media. Upscaling of the mean uniform transient flow in heterogeneous formations yields a time-dependent effective conductivity. As a result, the corresponding optimization of the aforementioned loss function is not trivial and, therefore, a flow simulator is developed with machine-learning software that uses state-of-the-art optimization methods. This methodology provides results that coincide with results obtained with traditional methods and further expands them in two major ways. First, it yields a probabilistic distribution of the upscaled parameter rather than a single value (the ensemble mean); this allows one to quantify prediction uncertainty of the upscaling procedure. Second, our methodology does not impose any physical constraints and limitations, such as the assumption of mild heterogeneity of a porous medium that underpins the perturbation-based strategies for conductivity scaling. In addition, our information-theoretic methodology is further expanded to tackle multi-dimensional transient flows. The change of direction of the mean flow induces anisotropy in the upscaled conductivity tensor, whose full effects cannot be captured with a two-point flux-approximation simulator. For this reason, a multi-point flux-approximation simulator is developed using the machine learning software. This allows us to predict the temporal evolution of all the components of the upscaled conductivity tensor. Finally, this study examines the use of neural networks as a surrogate model for Markov chain Monte Carlo (MCMC). For this purpose, a two-dimensional encoder-decoder convolutional neural network (CNN), which has been previously developed for inverse problems, is modified. The specific application considered in this thesis is a thermal-hydrologic-chemical model. The main challenges that arise in this setting are the large number of inputs and outputs needed to describe the problem. We show that the CNN surrogate is able to capture all the complex physical processes and is accurate enough to be used for MCMC. In addition, since all the gradient information is known, it enables the use of more evolved MCMC algorithms such as Hamiltonian Monte Carlo.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798357512581Subjects--Topical Terms:

619355
Software.
Index Terms--Genre/Form:

542853
Electronic books.

Information-Theoretic Approach for Upscaling.
LDR:04664nmm a2200397K 4500 001 2362644
005 20231102122746.5
006 m o d
007 cr mn ---uuuuu
008 241011s2022 xx obm 000 0 eng d
020 $a 9798357512581
035 $a (MiAaPQ)AAI29756186
035 $a (MiAaPQ)STANFORDmp002vz0000
035 $a AAI29756186
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Belivanis, Dimitrios Ioannis. $3 3703389
245 1 0 $a Information-Theoretic Approach for Upscaling.
264 0 $c 2022
300 $a 1 online resource (108 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 84-05, Section: A.
500 $a Advisor: Ahmmed, Bulbul;Mukerji, Tapan;Tartakovsky, Daniel.
502 $a Thesis (Ph.D.)--Stanford University, 2022.
504 $a Includes bibliographical references
520 $a Recent advances in hardware, algorithms, and scientific computing open new possibilities to revisit long-standing problems with additional tools in our arsenal. Examples of such problems, which are explored in this study, are inverse problems and upscaling of dynamic models. Both problems include nonlinear and functional minimization of the discrepancy between data and model predictions. This study develops an information-theoretic approach for upscaling of dynamic models affected by uncertainty. An optimal, time-dependent, probabilistic characterization of the macroscale model is obtained to yield minimum discrepancy with respect to assigned quantities of interest provided by the microscale solution. Hard data forming a training set are obtained at the fine scale from repeated synthetic simulations or from observations. Fine-scale information is transferred to the coarse scale via minimization of a loss function that consists of the cumulative average discrepancy, regularized by the cumulative exchange of information measured via mutual information. This general procedure is applied to transient flow processes in heterogeneous media. Upscaling of the mean uniform transient flow in heterogeneous formations yields a time-dependent effective conductivity. As a result, the corresponding optimization of the aforementioned loss function is not trivial and, therefore, a flow simulator is developed with machine-learning software that uses state-of-the-art optimization methods. This methodology provides results that coincide with results obtained with traditional methods and further expands them in two major ways. First, it yields a probabilistic distribution of the upscaled parameter rather than a single value (the ensemble mean); this allows one to quantify prediction uncertainty of the upscaling procedure. Second, our methodology does not impose any physical constraints and limitations, such as the assumption of mild heterogeneity of a porous medium that underpins the perturbation-based strategies for conductivity scaling. In addition, our information-theoretic methodology is further expanded to tackle multi-dimensional transient flows. The change of direction of the mean flow induces anisotropy in the upscaled conductivity tensor, whose full effects cannot be captured with a two-point flux-approximation simulator. For this reason, a multi-point flux-approximation simulator is developed using the machine learning software. This allows us to predict the temporal evolution of all the components of the upscaled conductivity tensor. Finally, this study examines the use of neural networks as a surrogate model for Markov chain Monte Carlo (MCMC). For this purpose, a two-dimensional encoder-decoder convolutional neural network (CNN), which has been previously developed for inverse problems, is modified. The specific application considered in this thesis is a thermal-hydrologic-chemical model. The main challenges that arise in this setting are the large number of inputs and outputs needed to describe the problem. We show that the CNN surrogate is able to capture all the complex physical processes and is accurate enough to be used for MCMC. In addition, since all the gradient information is known, it enables the use of more evolved MCMC algorithms such as Hamiltonian Monte Carlo.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Software. $2 gtt. $3 619355
650 4 $a Deep learning. $3 3554982
650 4 $a Histograms. $3 3683681
650 4 $a Inverse problems. $3 3686950
650 4 $a Permeability. $3 915594
650 4 $a Normal distribution. $3 3561025
650 4 $a Optimization. $3 891104
650 4 $a Neural networks. $3 677449
650 4 $a Data assimilation. $3 3682570
650 4 $a Homogenization. $3 3561160
650 4 $a Algorithms. $3 536374
650 4 $a Information theory. $3 542527
650 4 $a Expected values. $3 3563993
650 4 $a Hydraulics. $3 669758
650 4 $a Markov analysis. $3 3562906
650 4 $a Parameter estimation. $3 567557
650 4 $a Applied mathematics. $3 2122814
650 4 $a Computer science. $3 523869
650 4 $a Hydraulic engineering. $3 613609
650 4 $a Mathematics. $3 515831
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0364
690 $a 0800
690 $a 0984
690 $a 0218
690 $a 0338
690 $a 0405
690 $a 0796
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a Stanford University. $3 754827
773 0 $t Dissertations Abstracts International $g 84-05A.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29756186 $z click for full text (PQDT)