Mohammad , profile picture
Uploaded byMohammad
166 views

Development of Multi-Level ROM

This document describes a multi-level reduced order modeling approach with robust error bounds. It discusses applying dimensionality reduction algorithms to extract active subspaces from reduced complexity models, then equipping the reduced model with an error bound. A case study applies this approach to a nuclear reactor assembly model by extracting active subspaces from individual pin cell models to build a reduced order model in a more computationally efficient way than using the full assembly model.

Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Multi-level Reduced Order Modeling with Robust Error Bounds Mohammad G. Abdo and Hany S. Abdel-Khalik School of Nuclear Engineering Purdue University mgabdo@ncsu.edu and abdelkhalik@purdue.edu June 30, 2015 1 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Motivation ROM is indispensible for analysis with repetitive executions. ROM premised on the assumption: intrinsic dimensionality nominal dimensionality. ROM discards componants with negligible impact on reactor attributes of interest and hence must be equipped with error metrics. Can extract "active subspaces" from a reduced complexity model that undergoes similar physics. 2 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Objectives Apply the reduction and identify the active subspaces in a much more efficient methodology. Equip the reduced model with a robust error bound that can test the representitaveness of the active subspaces and hence define a validation domain that includes different conditions and different scenarios. 3 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Definition A nonlinear function f is said to be reducable if there exist matrices Urx ∈ Rn×rx and/or Ury ∈ Rm×ry such that: f(x) − Ury UT ry f(Urx UT rx x) f(x) ≤ 4 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Reduction Algorithms In our context, reduction algorithms refer to two different algorithms [4], each is used at a different interface: Gradient-free Snapshot Reduction Algorithm (Reduces response interface). Gradient-based Reduction Algorithm(Reduces parameter interface). 5 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 6 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 2 Execute the forward model k times and record the corresponding k responses: φi = f xi k i=1 , and aggregate them into: = φ1 φ2 · · · φk ∈ Rm×k . 6 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 2 Execute the forward model k times and record the corresponding k responses: φi = f xi k i=1 , and aggregate them into: = φ1 φ2 · · · φk ∈ Rm×k . 6 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Snapshot Reduction (cont.) 3 Calculate the singular value decomposition (SVD): = Uy Sy VT y ; where Uy ∈ Rm×k . 4 Collect the first ry columns of Uy in Ury to span the active response subspace, where ry ≤ min (m, k). 7 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Snapshot Reduction (cont.) 3 Calculate the singular value decomposition (SVD): = Uy Sy VT y ; where Uy ∈ Rm×k . 4 Collect the first ry columns of Uy in Ury to span the active response subspace, where ry ≤ min (m, k). 7 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-based Reduction This algorithm may be described by the following steps: 1 Execute the adjoint model k times to get: G = dR pseudo 1 dx x1 · · · dR pseudo k dx xk . 8 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-based Reduction This algorithm may be described by the following steps: 1 Execute the adjoint model k times to get: G = dR pseudo 1 dx x1 · · · dR pseudo k dx xk . 2 From SVD of: G = Ux Sx VT x , one can pick the first rx columns of Ux (denoted by Urx ) to span the active parameter subspace. 8 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation To estimate the error resulting from the reduction, the ijth entry of the operator E can be written as: [E]ij = fi xj − Ury (i, :) UT ry (i, :) fi Urx UT rx xj fi xj , where Urx ∈ Rn×rx and Ury ∈ Rm×ry are matrices whose orthonormal columns span the parameter and response spaces respectively. We need to estimate an upper bound for the error in each response. 9 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation (cont.) The 2-norm of E (or each row in E) can be estimated using: P E ≤ η max i=1,2,...s Ew(i) ≥ 1 − 1 η2 0 pdfw2 1 (t) dt s (2) where E ∈ Rm×N with N being the number of sampled responses and w is an N-dimensional random vector sampled from a known distribution D This probabilistic statement has its roots in Dixon’s theory [1983], where he sampled w(i) from a standard normal distribution and found an analytic value for the probability in terms of the multiplier η It is intuitive that if the user presets a probability of sucess then the value of the multiplier η depends solely on D from which w is sampled. careful inspection showed that the estimated error can be multiple order of magnitudes larger than the actual error (unneccessarily conservative). 10 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation (cont.) This motivated the numerical inspection of many distributions and the selection of the most practical one (i.e. which gave the least multiplier η). The inspection showed that the distribution which gave the least η is the binomial distribution. [2, 3] 11 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Distribution Selection Figure : Uniform Distribution. Figure : Gaussian Distribution. Estimated norm is orders of magnitude off. 12 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Distribution Selection [cont.] The binomial shows a linear structure arround the 45-degree solid line. This means that even if the case is a failure case (i.e. The actual norm is greater than the bound),the estimated norm will still be very close to the actual norm This appealing behaviour motivates the use of the binomial distribution to get rid of unneccessarily conservative bounds. Figure : Binomial Distribution 13 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 Benchmark lattice for Peach Bottom Atomic Power Station Unit2 (PB-2, 1112MWe BWR designed by OECD/NEA and manufactured by General Electric). Figure : 7x7 BWR Benchmark. 14 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . 15 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . Can we extract the active subspace from running only subdomain of the problem? Or a reduced complexity model? 15 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . Can we extract the active subspace from running only subdomain of the problem? Or a reduced complexity model? Figure : Calculation Levels. http://www.nrc.gov/about-nrc/emerg-preparedness/images/fuel-pellet- assembly.jpg 15 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. 16 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! 16 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! 16 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 16 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 17 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 18 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 19 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 20 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 21 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 22 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 23 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 24 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 25 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. 26 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. This motivates that the active parameter subspace for the whole assembly might be revealed from sampling a pin or more !! 26 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. This motivates that the active parameter subspace for the whole assembly might be revealed from sampling a pin or more !! Tests Description: Identify the parameter active subspace for one (or more) pin cells Construct an error bound for each response test the identified subspace on different pins then on the whole assembly. If successful, test it in different conditions !! 26 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] Nominal dimension for the parameter space n = 127nuclides ×7reactions ×56 energy groups = 49784 3 pins are depleted to 30 GWd/MTU then used to extract the subspace (2.93% UO2 with 3 % gd, 1.94%UO2,2.93%UO2 ) and rx is taken to be 1500 Test1: The subspace is tested on the highest enrichment pin and a completely different one. First two figures will show the error in flux within two selective energy ranges: 1.85-3.0 MeV and 0.625-1.01 eV for the most dominant Pin Cell. Then a figure showing the maximum and mean errors over all energies and hence are taken as the bounds. Test2: This is repeated for another pin cell (Mixture 4) Test3: The extracted subspace is then employed on the whole assembly depleted to the same point (30GWd/MTU). The results for this test is shown in 9 figures showing the maximum and mean error for the 9 unique mixtures. 27 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 1 Figure : Fast Flux Error (Mix 500, LF). Figure : Thermal Flux Error(Mix 500, LF). 28 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 2 Figure : Error Boubds (Mix 500, LF). Figure : Actual Errors(Mix 4, LF). The left figure shows the typical bounds doesn’t exceed 3% for maximum error and is less than 0.7% for mean error. 29 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 30 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 31 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 32 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 33 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 34 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 The assembly is depleted to (60 GWd/MTU) The next 9 figures show the maximum and mean errors for the 9 different mixtures. This test aims to inspect the behaviour of the active subspace extracted from the low fidelity model at different composition due to different burnup. 35 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 36 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 37 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 38 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 39 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 40 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 The assembly is depleted to the End of the first cycle (20 GWd/MTU) and at Cold conditions The next 9 figures show the maximum and mean errors for the 9 different mixtures. This test aims to inspect the behaviour of the active subspace extracted from the low fidelity model at different composition due to different burnup and at different temperature. 41 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 42 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 43 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 44 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 45 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 46 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Conclusions ROM errors are reliably quantified using realistic bounds (i.e., actual error is close to error bound). Provide-a-first-of-a-kind approach to quantify errors resulting from dimensionality reduction in nuclear reactor calculations. Can be used to experiment with different ROM techniques to determine optimum performance for application of interest. Quantify errors resulting from homogenization theory (a form of dimensionality reduction). Multi-physics ROM, where one physics determines active subspace for next physics. Efficient rendering of active subspaces for expensive model, MLROM 47 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Conclusions ROM errors are reliably quantified using realistic bounds (i.e., actual error is close to error bound). Provide-a-first-of-a-kind approach to quantify errors resulting from dimensionality reduction in nuclear reactor calculations. Can be used to experiment with different ROM techniques to determine optimum performance for application of interest. Quantify errors resulting from homogenization theory (a form of dimensionality reduction). Multi-physics ROM, where one physics determines active subspace for next physics. Efficient rendering of active subspaces for expensive model, MLROM Using MLROM enables the application of ROM on all models that was expensive enough to prohebit executing the model to extract the subspace. This enables the determination of the validation space and One can make a statement about how good is the subspace if used with conditions different than those used in the sampling process. 47 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Acknowledgements I’d like to acknowledge the support of the Department of Nuclear Engineering at North Carolina State University to complete this work in support of my PhD. 48 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Bibliography I SCALE:A Comperhensive Modeling and Simulation Suite for Nuclear Safety Analysis and Design,ORNL/TM-2005/39, Version 6.1, Oak Ridge National Laboratory, Oak Ridge, Tennessee,June 2011. Available from Radiation Safety Information Computational Center at Oak Rodge National Laboratory as CCC-785. M. G. ABDO AND H. S. ABDEL-KHALIK, Propagation of error bounds due to active subspace reduction, ANS, 110 (2014), pp. 196–199. , Further investigation of error bounds for reduced order modeling, ANS MC2015, (2015). Y. BANG, J. HITE, AND H. S. ABDEL-KHALIK, Hybrid reduced order modeling applied to nonlinear models, IJNME, 91 (2012), pp. 929–949. J. D. DIXON, Estimating extremal eigenvalues and condition numbers of matrices, SIAM, 20 (1983), pp. 812–814. 49 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Bibliography II N. HALKO, P. G. MARTINSSON, AND J. A. TROPP, Finding structure with randomness:probabilistic algorithms for constructing approximate matrix decompositions, SIAM, 53 (2011), pp. 217–288. P. G. MARTINSSON, V. ROKHLIN, AND M. TYGERT, A randomized algorithm for the approximation of matrices, tech. report, Yale University. J. A. TROPP, User-friendly tools for random matrices. S. S. WILKS, Mathematical statistics, John Wiley, New York, 1st ed., 1962. F. WOOLFE, E. LIBERTY, V. ROKHLIN, AND M. TYGERT, A fast randomized algorithm for the approximation of matrices, preliminary report, Yale University. 50 / 51
Motivation Objectives Background of Supporting Algorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Questions/Suggestions? 51 / 51

More Related Content

PDF
Static Analysis and Verification of C Programs
byNew York City College of Technology Computer Systems Technology Colloquium
 
PPTX
Optimization problems and algorithms
byAboul Ella Hassanien
 
PDF
ANSSummer2015
byMohammad Abdo
 
PPT
Chapter 10 ds
byHanif Durad
 
PDF
Training and Inference for Deep Gaussian Processes
byKeyon Vafa
 
DOC
352735340 rsh-qam11-tif-12-doc
byFiras Husseini
 
DOC
352735339 rsh-qam11-tif-13-doc
byFiras Husseini
 
PDF
MULTI-OBJECTIVE ENERGY EFFICIENT OPTIMIZATION ALGORITHM FOR COVERAGE CONTROL ...
byijcseit
 
Static Analysis and Verification of C Programs
byNew York City College of Technology Computer Systems Technology Colloquium
 
Optimization problems and algorithms
byAboul Ella Hassanien
 
ANSSummer2015
byMohammad Abdo
 
Chapter 10 ds
byHanif Durad
 
Training and Inference for Deep Gaussian Processes
byKeyon Vafa
 
352735340 rsh-qam11-tif-12-doc
byFiras Husseini
 
352735339 rsh-qam11-tif-13-doc
byFiras Husseini
 
MULTI-OBJECTIVE ENERGY EFFICIENT OPTIMIZATION ALGORITHM FOR COVERAGE CONTROL ...
byijcseit
 

What's hot

PPTX
Unit 5
byGunasundari Selvaraj
 
PPT
Instance Based Learning in Machine Learning
byPavithra Thippanaik
 
PDF
Chap 8. Optimization for training deep models
byYoung-Geun Choi
 
PDF
17 Machine Learning Radial Basis Functions
byAndres Mendez-Vazquez
 
PPTX
Instance based learning
byswapnac12
 
PDF
Safety Verification of Deep Neural Networks_.pdf
bylarbaoui
 
PPT
Chapter 3 pc
byHanif Durad
 
PDF
Adversarial Reinforced Learning for Unsupervised Domain Adaptation
bytaeseon ryu
 
PDF
Backtracking & branch and bound
byVipul Chauhan
 
PPT
Chapter 2 ds
byHanif Durad
 
PDF
QUESTION BANK FOR ANNA UNNIVERISTY SYLLABUS
byJAMBIKA
 
PDF
Mining at scale with latent factor models for matrix completion
byFabio Petroni, PhD
 
PDF
Daa notes 2
bysmruti sarangi
 
PDF
Polynomial Matrix Decompositions
byFörderverein Technische Fakultät
 
PDF
Statistical Pattern recognition(1)
bySyed Atif Naseem
 
PDF
design and analysis of algorithm
byMuhammad Arish
 
PDF
main
byAbigail See
 
PDF
論文紹介 Combining Model-Based and Model-Free Updates for Trajectory-Centric Rein...
byKusano Hitoshi
 
PDF
LCBM: Statistics-Based Parallel Collaborative Filtering
byFabio Petroni, PhD
 
PDF
Analysing and combining partial problem solutions for properly informed heuri...
byAlexander Decker
 
Unit 5
byGunasundari Selvaraj
 
Instance Based Learning in Machine Learning
byPavithra Thippanaik
 
Chap 8. Optimization for training deep models
byYoung-Geun Choi
 
17 Machine Learning Radial Basis Functions
byAndres Mendez-Vazquez
 
Instance based learning
byswapnac12
 
Safety Verification of Deep Neural Networks_.pdf
bylarbaoui
 
Chapter 3 pc
byHanif Durad
 
Adversarial Reinforced Learning for Unsupervised Domain Adaptation
bytaeseon ryu
 
Backtracking & branch and bound
byVipul Chauhan
 
Chapter 2 ds
byHanif Durad
 
QUESTION BANK FOR ANNA UNNIVERISTY SYLLABUS
byJAMBIKA
 
Mining at scale with latent factor models for matrix completion
byFabio Petroni, PhD
 
Daa notes 2
bysmruti sarangi
 
Polynomial Matrix Decompositions
byFörderverein Technische Fakultät
 
Statistical Pattern recognition(1)
bySyed Atif Naseem
 
design and analysis of algorithm
byMuhammad Arish
 
main
byAbigail See
 
論文紹介 Combining Model-Based and Model-Free Updates for Trajectory-Centric Rein...
byKusano Hitoshi
 
LCBM: Statistics-Based Parallel Collaborative Filtering
byFabio Petroni, PhD
 
Analysing and combining partial problem solutions for properly informed heuri...
byAlexander Decker
 

Viewers also liked

PPT
MIM Architecture Overview
byTim Dowty
 
PDF
Bourntec, INC
byS Reddy Veerareddy
 
PPT
Coal /coal ash/tar sands=GLOBAL WARMING.........
byOpalcreek
 
PPTX
Zoyo Neighborhood Yogurt Franchise in Michigan
byEmily Verbeke
 
PDF
Launching Smart Use of Medicine Campaign
byMay Haddad MD.MPH
 
PDF
CHECKING ACTIONS IN DCPDS
byDaniela Stokes
 
PDF
CA Varun Sethi - IndAS 102 - IFRS 2 - Share based payments - Accounting for ...
byVarun Sethi
 
DOC
Dr. elham cv 2013 final3 (1)
byElham Aziz
 
DOCX
Tugas switching
byRizky d'Cyber
 
DOC
Mike Williams
byMike Williams
 
DOC
Cv elham programe الامارات نهائى5 2015
byElham Aziz
 
DOC
FRM220_W2_A1
byLinnie Williams Garrett
 
PPTX
Stock market lm
byGKTheodora
 
PDF
Listing Presentation
byAlexander De Luna
 
PDF
Acer aspire v5 122 p-4215..
bypatrickkayebula
 
PDF
Tercera parte estatuto tributario
byangelik2015
 
PPTX
Beacon hill catering
byKarenLSeitz
 
DOCX
IRE2
byCollby Soo
 
PPTX
ΑΡΚΟΥΔΑΚΙ
bydimotiko_delasalle
 
MIM Architecture Overview
byTim Dowty
 
Bourntec, INC
byS Reddy Veerareddy
 
Coal /coal ash/tar sands=GLOBAL WARMING.........
byOpalcreek
 
Zoyo Neighborhood Yogurt Franchise in Michigan
byEmily Verbeke
 
Launching Smart Use of Medicine Campaign
byMay Haddad MD.MPH
 
CHECKING ACTIONS IN DCPDS
byDaniela Stokes
 
CA Varun Sethi - IndAS 102 - IFRS 2 - Share based payments - Accounting for ...
byVarun Sethi
 
Dr. elham cv 2013 final3 (1)
byElham Aziz
 
Tugas switching
byRizky d'Cyber
 
Mike Williams
byMike Williams
 
Cv elham programe الامارات نهائى5 2015
byElham Aziz
 
FRM220_W2_A1
byLinnie Williams Garrett
 
Stock market lm
byGKTheodora
 
Listing Presentation
byAlexander De Luna
 
Acer aspire v5 122 p-4215..
bypatrickkayebula
 
Tercera parte estatuto tributario
byangelik2015
 
Beacon hill catering
byKarenLSeitz
 
IRE2
byCollby Soo
 
ΑΡΚΟΥΔΑΚΙ
bydimotiko_delasalle
 

Similar to Development of Multi-Level ROM

PDF
Propagation of Error Bounds due to Active Subspace Reduction
byMohammad
 
PDF
Development of Multi-level Reduced Order MOdeling Methodology
byMohammad
 
PDF
Workshop in honour of Don Poskitt and Gael Martin
byChristian Robert
 
PDF
Laplace's Demon: seminar #1
byChristian Robert
 
PDF
Numerical Solution of Linear algebraic Equation
bypayalpriyadarshinisa1
 
PPT
Cs221 lecture4-fall11
bydarwinrlo
 
PDF
AbdoSummerANS_mod3
byMohammad Abdo
 
PDF
Prpagation of Error Bounds Across reduction interfaces
byMohammad
 
PDF
ProbErrorBoundROM_MC2015
byMohammad Abdo
 
PDF
Probabilistic Error Bounds for Reduced Order Modeling M&C2015
byMohammad
 
PDF
FurtherInvestegationOnProbabilisticErrorBounds_final
byMohammad Abdo
 
PDF
Further investegationonprobabilisticerrorbounds final
byMohammad
 
PDF
MultiLevelROM_ANS_Summer2015_RevMarch23
byMohammad Abdo
 
PDF
MC2015Posterlandscape
byMohammad Abdo
 
PDF
Mc2015 posterlandscape
byMohammad
 
PDF
ASQ Talk v4
byGhulam Mustafa, Ph. D.
 
PDF
the ABC of ABC
byChristian Robert
 
PDF
asymptotics of ABC
byChristian Robert
 
PDF
QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...
byThe Statistical and Applied Mathematical Sciences Institute
 
PDF
Asymptotics of ABC, lecture, Collège de France
byChristian Robert
 
Propagation of Error Bounds due to Active Subspace Reduction
byMohammad
 
Development of Multi-level Reduced Order MOdeling Methodology
byMohammad
 
Workshop in honour of Don Poskitt and Gael Martin
byChristian Robert
 
Laplace's Demon: seminar #1
byChristian Robert
 
Numerical Solution of Linear algebraic Equation
bypayalpriyadarshinisa1
 
Cs221 lecture4-fall11
bydarwinrlo
 
AbdoSummerANS_mod3
byMohammad Abdo
 
Prpagation of Error Bounds Across reduction interfaces
byMohammad
 
ProbErrorBoundROM_MC2015
byMohammad Abdo
 
Probabilistic Error Bounds for Reduced Order Modeling M&C2015
byMohammad
 
FurtherInvestegationOnProbabilisticErrorBounds_final
byMohammad Abdo
 
Further investegationonprobabilisticerrorbounds final
byMohammad
 
MultiLevelROM_ANS_Summer2015_RevMarch23
byMohammad Abdo
 
MC2015Posterlandscape
byMohammad Abdo
 
Mc2015 posterlandscape
byMohammad
 
ASQ Talk v4
byGhulam Mustafa, Ph. D.
 
the ABC of ABC
byChristian Robert
 
asymptotics of ABC
byChristian Robert
 
QMC: Transition Workshop - Probability Models for Discretization Uncertainty ...
byThe Statistical and Applied Mathematical Sciences Institute
 
Asymptotics of ABC, lecture, Collège de France
byChristian Robert
 

Recently uploaded

PDF
Lab06-Containerized applications deployment and management using Kubernetes.pdf
byKhucBaoMinhB2105709
 
PDF
JWT_Authentication_MERN_Stack_Assignment.pdf
byJaydeep Kale
 
PDF
OpenInfra Europe 2025: OVN traffic flow & Troubleshooting
byVadim Ponomarev
 
PPT
Chapter 5 Arithmetic Functions Computer Organisation and Architecture
byRitikSharma297999
 
PDF
Test Blueprint for National Exit Examination.pdf
byshumibiru238
 
PDF
Redefining the Sky with AI-Powered Autonomous Drones
byRiku Nagano
 
PDF
6.IPE431_KINEMATIC STRUCTURE2133e324.pdf
bygixaj71508
 
PDF
Impact of Different Curing Techniques in Evaluating the Strength of the Rolle...
byAbdullah Al-Ani
 
DOCX
Summer training report on plc and scada.docx
byharshgahlyan3108
 
PDF
2025 MediaEval Medico Organizer Presentation
bySushant Gautam
 
PDF
Step by Step Guide to OVERHAULING GENERATORS on Ships
byMahmoud Moghtaderi
 
PPTX
Green-ICT-and-Sustainable-Technology-Powering-a-Greener-Digital-Future.pptx
byRAJVEEPATEL13
 
PPTX
Pavement Quality Concrete Design Lecture.pptx
byAditya Rajan Patra
 
PPTX
Rahul Sharma.pptx for summer internshipp
byirfan7231089794
 
PPTX
Refrigeration and air conditioning .pptx
bykhushicapricorn1001
 
PDF
AI-DRIVEN CTI FOR BUSINESS: EMERGING THREATS, ATTACK STRATEGIES, AND DEFENSIV...
byAIRCC Publishing Corporation
 
PPTX
Trade Secret.pptx (Trade Secret.pptx.pptx)
byganeshchalla7
 
PDF
Resonating 25 years of service- 1988 exam batch.pdf
byRajesh Prasad
 
PDF
Meraki Cloud Management and Monitoring for Catalyst - Cisco Certificate
byVICTOR MAESTRE RAMIREZ
 
PDF
JLEE Electrical Engineering Question Paper Assam
bykandramariana6
 
Lab06-Containerized applications deployment and management using Kubernetes.pdf
byKhucBaoMinhB2105709
 
JWT_Authentication_MERN_Stack_Assignment.pdf
byJaydeep Kale
 
OpenInfra Europe 2025: OVN traffic flow & Troubleshooting
byVadim Ponomarev
 
Chapter 5 Arithmetic Functions Computer Organisation and Architecture
byRitikSharma297999
 
Test Blueprint for National Exit Examination.pdf
byshumibiru238
 
Redefining the Sky with AI-Powered Autonomous Drones
byRiku Nagano
 
6.IPE431_KINEMATIC STRUCTURE2133e324.pdf
bygixaj71508
 
Impact of Different Curing Techniques in Evaluating the Strength of the Rolle...
byAbdullah Al-Ani
 
Summer training report on plc and scada.docx
byharshgahlyan3108
 
2025 MediaEval Medico Organizer Presentation
bySushant Gautam
 
Step by Step Guide to OVERHAULING GENERATORS on Ships
byMahmoud Moghtaderi
 
Green-ICT-and-Sustainable-Technology-Powering-a-Greener-Digital-Future.pptx
byRAJVEEPATEL13
 
Pavement Quality Concrete Design Lecture.pptx
byAditya Rajan Patra
 
Rahul Sharma.pptx for summer internshipp
byirfan7231089794
 
Refrigeration and air conditioning .pptx
bykhushicapricorn1001
 
AI-DRIVEN CTI FOR BUSINESS: EMERGING THREATS, ATTACK STRATEGIES, AND DEFENSIV...
byAIRCC Publishing Corporation
 
Trade Secret.pptx (Trade Secret.pptx.pptx)
byganeshchalla7
 
Resonating 25 years of service- 1988 exam batch.pdf
byRajesh Prasad
 
Meraki Cloud Management and Monitoring for Catalyst - Cisco Certificate
byVICTOR MAESTRE RAMIREZ
 
JLEE Electrical Engineering Question Paper Assam
bykandramariana6
 

Development of Multi-Level ROM

  • 1.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Multi-level Reduced Order Modeling with Robust Error Bounds Mohammad G. Abdo and Hany S. Abdel-Khalik School of Nuclear Engineering Purdue University [email protected] and [email protected] June 30, 2015 1 / 51
  • 2.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Motivation ROM is indispensible for analysis with repetitive executions. ROM premised on the assumption: intrinsic dimensionality nominal dimensionality. ROM discards componants with negligible impact on reactor attributes of interest and hence must be equipped with error metrics. Can extract "active subspaces" from a reduced complexity model that undergoes similar physics. 2 / 51
  • 3.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Objectives Apply the reduction and identify the active subspaces in a much more efficient methodology. Equip the reduced model with a robust error bound that can test the representitaveness of the active subspaces and hence define a validation domain that includes different conditions and different scenarios. 3 / 51
  • 4.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Definition A nonlinear function f is said to be reducable if there exist matrices Urx ∈ Rn×rx and/or Ury ∈ Rm×ry such that: f(x) − Ury UT ry f(Urx UT rx x) f(x) ≤ 4 / 51
  • 5.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Reduction Algorithms In our context, reduction algorithms refer to two different algorithms [4], each is used at a different interface: Gradient-free Snapshot Reduction Algorithm (Reduces response interface). Gradient-based Reduction Algorithm(Reduces parameter interface). 5 / 51
  • 6.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 6 / 51
  • 7.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 2 Execute the forward model k times and record the corresponding k responses: φi = f xi k i=1 , and aggregate them into: = φ1 φ2 · · · φk ∈ Rm×k . 6 / 51
  • 8.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-free Snapshot Reduction Algorithm Consider the reducible model under inspection to be described by: φ = f (x) , (1) 1 Generate k random parameters realizations: xi k i=1. 2 Execute the forward model k times and record the corresponding k responses: φi = f xi k i=1 , and aggregate them into: = φ1 φ2 · · · φk ∈ Rm×k . 6 / 51
  • 9.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Snapshot Reduction (cont.) 3 Calculate the singular value decomposition (SVD): = Uy Sy VT y ; where Uy ∈ Rm×k . 4 Collect the first ry columns of Uy in Ury to span the active response subspace, where ry ≤ min (m, k). 7 / 51
  • 10.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Snapshot Reduction (cont.) 3 Calculate the singular value decomposition (SVD): = Uy Sy VT y ; where Uy ∈ Rm×k . 4 Collect the first ry columns of Uy in Ury to span the active response subspace, where ry ≤ min (m, k). 7 / 51
  • 11.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-based Reduction This algorithm may be described by the following steps: 1 Execute the adjoint model k times to get: G = dR pseudo 1 dx x1 · · · dR pseudo k dx xk . 8 / 51
  • 12.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Gradient-based Reduction This algorithm may be described by the following steps: 1 Execute the adjoint model k times to get: G = dR pseudo 1 dx x1 · · · dR pseudo k dx xk . 2 From SVD of: G = Ux Sx VT x , one can pick the first rx columns of Ux (denoted by Urx ) to span the active parameter subspace. 8 / 51
  • 13.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation To estimate the error resulting from the reduction, the ijth entry of the operator E can be written as: [E]ij = fi xj − Ury (i, :) UT ry (i, :) fi Urx UT rx xj fi xj , where Urx ∈ Rn×rx and Ury ∈ Rm×ry are matrices whose orthonormal columns span the parameter and response spaces respectively. We need to estimate an upper bound for the error in each response. 9 / 51
  • 14.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation (cont.) The 2-norm of E (or each row in E) can be estimated using: P E ≤ η max i=1,2,...s Ew(i) ≥ 1 − 1 η2 0 pdfw2 1 (t) dt s (2) where E ∈ Rm×N with N being the number of sampled responses and w is an N-dimensional random vector sampled from a known distribution D This probabilistic statement has its roots in Dixon’s theory [1983], where he sampled w(i) from a standard normal distribution and found an analytic value for the probability in terms of the multiplier η It is intuitive that if the user presets a probability of sucess then the value of the multiplier η depends solely on D from which w is sampled. careful inspection showed that the estimated error can be multiple order of magnitudes larger than the actual error (unneccessarily conservative). 10 / 51
  • 15.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Error Estimation (cont.) This motivated the numerical inspection of many distributions and the selection of the most practical one (i.e. which gave the least multiplier η). The inspection showed that the distribution which gave the least η is the binomial distribution. [2, 3] 11 / 51
  • 16.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Distribution Selection Figure : Uniform Distribution. Figure : Gaussian Distribution. Estimated norm is orders of magnitude off. 12 / 51
  • 17.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks ROM Algorithms Error Estimation Distribution Selection [cont.] The binomial shows a linear structure arround the 45-degree solid line. This means that even if the case is a failure case (i.e. The actual norm is greater than the bound),the estimated norm will still be very close to the actual norm This appealing behaviour motivates the use of the binomial distribution to get rid of unneccessarily conservative bounds. Figure : Binomial Distribution 13 / 51
  • 18.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 Benchmark lattice for Peach Bottom Atomic Power Station Unit2 (PB-2, 1112MWe BWR designed by OECD/NEA and manufactured by General Electric). Figure : 7x7 BWR Benchmark. 14 / 51
  • 19.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . 15 / 51
  • 20.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . Can we extract the active subspace from running only subdomain of the problem? Or a reduced complexity model? 15 / 51
  • 21.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] The idea is that for such an assembly, running the forward model and the assembly model rx times to identify the active parameter subspace is still very expensive !! . Can we extract the active subspace from running only subdomain of the problem? Or a reduced complexity model? Figure : Calculation Levels. http://www.nrc.gov/about-nrc/emerg-preparedness/images/fuel-pellet- assembly.jpg 15 / 51
  • 22.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. 16 / 51
  • 23.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! 16 / 51
  • 24.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! 16 / 51
  • 25.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 16 / 51
  • 26.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 17 / 51
  • 27.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 18 / 51
  • 28.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 19 / 51
  • 29.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 20 / 51
  • 30.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 21 / 51
  • 31.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 22 / 51
  • 32.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 23 / 51
  • 33.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 24 / 51
  • 34.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 1 [cont.] Why don’t we extract the parameters active subspace for each of the nine pin cells and try to visualize them ?!. (x ∈ R49784)!!!! We are not visualizing to see what does the topology of this n-dimensional surface look like ! We are only trying to see how far is each subspace from the other ! How different ! Figure : Scatter Visualization for the active subspaces. 25 / 51
  • 35.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. 26 / 51
  • 36.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. This motivates that the active parameter subspace for the whole assembly might be revealed from sampling a pin or more !! 26 / 51
  • 37.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] The previous figure defends that active parameter subspaces for the 9 pins are pretty close. This motivates that the active parameter subspace for the whole assembly might be revealed from sampling a pin or more !! Tests Description: Identify the parameter active subspace for one (or more) pin cells Construct an error bound for each response test the identified subspace on different pins then on the whole assembly. If successful, test it in different conditions !! 26 / 51
  • 38.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study [cont.] Nominal dimension for the parameter space n = 127nuclides ×7reactions ×56 energy groups = 49784 3 pins are depleted to 30 GWd/MTU then used to extract the subspace (2.93% UO2 with 3 % gd, 1.94%UO2,2.93%UO2 ) and rx is taken to be 1500 Test1: The subspace is tested on the highest enrichment pin and a completely different one. First two figures will show the error in flux within two selective energy ranges: 1.85-3.0 MeV and 0.625-1.01 eV for the most dominant Pin Cell. Then a figure showing the maximum and mean errors over all energies and hence are taken as the bounds. Test2: This is repeated for another pin cell (Mixture 4) Test3: The extracted subspace is then employed on the whole assembly depleted to the same point (30GWd/MTU). The results for this test is shown in 9 figures showing the maximum and mean error for the 9 unique mixtures. 27 / 51
  • 39.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 1 Figure : Fast Flux Error (Mix 500, LF). Figure : Thermal Flux Error(Mix 500, LF). 28 / 51
  • 40.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 2 Figure : Error Boubds (Mix 500, LF). Figure : Actual Errors(Mix 4, LF). The left figure shows the typical bounds doesn’t exceed 3% for maximum error and is less than 0.7% for mean error. 29 / 51
  • 41.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 30 / 51
  • 42.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 31 / 51
  • 43.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 32 / 51
  • 44.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 33 / 51
  • 45.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Test 3 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 34 / 51
  • 46.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 The assembly is depleted to (60 GWd/MTU) The next 9 figures show the maximum and mean errors for the 9 different mixtures. This test aims to inspect the behaviour of the active subspace extracted from the low fidelity model at different composition due to different burnup. 35 / 51
  • 47.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 36 / 51
  • 48.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 37 / 51
  • 49.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 38 / 51
  • 50.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 39 / 51
  • 51.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 40 / 51
  • 52.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 The assembly is depleted to the End of the first cycle (20 GWd/MTU) and at Cold conditions The next 9 figures show the maximum and mean errors for the 9 different mixtures. This test aims to inspect the behaviour of the active subspace extracted from the low fidelity model at different composition due to different burnup and at different temperature. 41 / 51
  • 53.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 1, HF). Figure : Actual Errors(Mix 2, HF). 42 / 51
  • 54.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 4, HF). Figure : Actual Errors(Mix 500, HF). 43 / 51
  • 55.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix 201, HF). Figure : Actual Errors(Mix 202, HF). 44 / 51
  • 56.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 3 [cont.] Figure : Actual Errors (Mix203, HF). Figure : Actual Errors(Mix 212, HF). 45 / 51
  • 57.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Case Study 1 Case Study 2 Case Study 3 Case Study 2 [cont.] Figure : Actual Errors (Mix 213, HF). Figures show that the subspace extracted from the low-fidelity model fully represent the full assembly at 30 GWd/MTU and hot conditions. the maximum error did not exceed 3% and the mean error is always less than 0.7% which is exactly consistent with what we had from the low-fidelity model. 46 / 51
  • 58.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Conclusions ROM errors are reliably quantified using realistic bounds (i.e., actual error is close to error bound). Provide-a-first-of-a-kind approach to quantify errors resulting from dimensionality reduction in nuclear reactor calculations. Can be used to experiment with different ROM techniques to determine optimum performance for application of interest. Quantify errors resulting from homogenization theory (a form of dimensionality reduction). Multi-physics ROM, where one physics determines active subspace for next physics. Efficient rendering of active subspaces for expensive model, MLROM 47 / 51
  • 59.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Conclusions ROM errors are reliably quantified using realistic bounds (i.e., actual error is close to error bound). Provide-a-first-of-a-kind approach to quantify errors resulting from dimensionality reduction in nuclear reactor calculations. Can be used to experiment with different ROM techniques to determine optimum performance for application of interest. Quantify errors resulting from homogenization theory (a form of dimensionality reduction). Multi-physics ROM, where one physics determines active subspace for next physics. Efficient rendering of active subspaces for expensive model, MLROM Using MLROM enables the application of ROM on all models that was expensive enough to prohebit executing the model to extract the subspace. This enables the determination of the validation space and One can make a statement about how good is the subspace if used with conditions different than those used in the sampling process. 47 / 51
  • 60.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Acknowledgements I’d like to acknowledge the support of the Department of Nuclear Engineering at North Carolina State University to complete this work in support of my PhD. 48 / 51
  • 61.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Bibliography I SCALE:A Comperhensive Modeling and Simulation Suite for Nuclear Safety Analysis and Design,ORNL/TM-2005/39, Version 6.1, Oak Ridge National Laboratory, Oak Ridge, Tennessee,June 2011. Available from Radiation Safety Information Computational Center at Oak Rodge National Laboratory as CCC-785. M. G. ABDO AND H. S. ABDEL-KHALIK, Propagation of error bounds due to active subspace reduction, ANS, 110 (2014), pp. 196–199. , Further investigation of error bounds for reduced order modeling, ANS MC2015, (2015). Y. BANG, J. HITE, AND H. S. ABDEL-KHALIK, Hybrid reduced order modeling applied to nonlinear models, IJNME, 91 (2012), pp. 929–949. J. D. DIXON, Estimating extremal eigenvalues and condition numbers of matrices, SIAM, 20 (1983), pp. 812–814. 49 / 51
  • 62.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Bibliography II N. HALKO, P. G. MARTINSSON, AND J. A. TROPP, Finding structure with randomness:probabilistic algorithms for constructing approximate matrix decompositions, SIAM, 53 (2011), pp. 217–288. P. G. MARTINSSON, V. ROKHLIN, AND M. TYGERT, A randomized algorithm for the approximation of matrices, tech. report, Yale University. J. A. TROPP, User-friendly tools for random matrices. S. S. WILKS, Mathematical statistics, John Wiley, New York, 1st ed., 1962. F. WOOLFE, E. LIBERTY, V. ROKHLIN, AND M. TYGERT, A fast randomized algorithm for the approximation of matrices, preliminary report, Yale University. 50 / 51
  • 63.
    Motivation Objectives Background of SupportingAlgorithms and Theory Numerical tests and results Conclusions Acknowledgements Bibliography Thanks Questions/Suggestions? 51 / 51