Download - Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Transcript
Page 1: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

8 mm

215

mm

Nested Multiscale Simulations

10

๐ฃ๐ฃ๐ญ๐ญ๐ญ๐ญ IP๐ข๐ข๐ญ๐ญ๐ญ๐ญ IP

๐ฃ๐ฃ๐ญ๐ญ๐ญ๐ญ IP

* Resin not shown

640,000 IPs

15,000 IPs

900,000 IPs

๐ข๐ข๐ญ๐ญ๐ญ๐ญ IP

๐ฃ๐ฃ๐ญ๐ญ๐ญ๐ญ IP

๐ข๐ข๐ญ๐ญ๐ญ๐ญ IP

10 ฮผm

80,000 IPs

โ€ข Random process based UQ in a single length-scaleโ€ข Neglecting spatial & part-to-part variations

Prior Works:

โ€ข Integration points (IPs) arestructures at finer scales.

๐›ผ๐›ผ = 90ยฐvf = 62%

๐›ผ๐›ผ = 80ยฐvf = 68%

vf = 65%

vf = 65%vf = 60%

vf = 70%vf = 64%

๐›ผ๐›ผ: Yarn angle

vf: Avg. fiber volume fraction

(3) Cured UD-RVE at Microscale

(2) Cured Woven-RVE at Mesoscale

(1) Woven Composite at Macroscale

Page 2: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Our Approach For UQ & UP in Multiscale Simulations

11

Meso StructureMeso SRPMeso Simulator

Macro StructureMacro SRPMacro Simulator

โ€ข Modeling uncertainties sources and their spatial variations via spatial random processes (SRPs)

โ€ข Coupling SRPs across length-scales via top-down sampling

โ€ข Creating metamodels of homogenized constitutive relations w microstructure variations

Bostanabad, R., et. al. (2018) CMAME, 338, 506-532.

Managing Complexity โ€“On-the-fly Machine Learning

Metamodeling

Homogenization

Homogenization Speed-up

Sensitivity Analysis

Input Variation

Out

put V

aria

tion

Micro Structure

Micro SRPMicro Simulator

Homogenization Top-down Sampling

Top-down Sampling

Sampling Dimension Reduction

Page 3: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

๐‘ฅ๐‘ฅ1

๐‘ฅ๐‘ฅ2

Multi-Response Gaussian Processes (MRGPs) for Quantifying Correlated Uncertain Sources

12

Multi-response Gaussian process (MRGP):

GP

MRGP

Only need mean and covariance functions.

Single response Gaussian processes:

๐’š๐’š~๐บ๐บ๐บ๐บ ๐‘š๐‘š ๏ฟฝ ,๐œŽ๐œŽ2๐’“๐’“ ๏ฟฝ,๏ฟฝ

๐’€๐’€~๐‘€๐‘€๐‘€๐‘€๐บ๐บ๐บ๐บ ๐’Ž๐’Ž ๏ฟฝ ,๐œฎ๐œฎโจ‚๐’“๐’“ ๏ฟฝ,๏ฟฝ

Mean Function Covariance Function

๐‘ฅ๐‘ฅ1

๐‘ฃ๐‘ฃ๐‘’๐‘’

๐œ™๐œ™

๐‘ฅ๐‘ฅ1

๐‘ฃ๐‘ฃ๐‘’๐‘’

โ€ข Captures the between-response correlations &variations around themeans.

๐œฎ๐œฎ = 4 โˆ’10โˆ’10 50

๐œŽ๐œŽ = 4,๐œ”๐œ” = 1.5,๐‘š๐‘š = 55GP:

MRGP:

๐œ”๐œ” = 1.5,๐‘š๐‘š = 550

๐‘Ÿ๐‘Ÿ ๐’™๐’™,๐’™๐’™โ€ฒ = ๐‘’๐‘’โˆ’ โˆ‘๐‘–๐‘–=1๐‘๐‘ 10๐œ”๐œ”๐‘–๐‘– ๐‘ฅ๐‘ฅ๐‘–๐‘–โˆ’๐‘ฅ๐‘ฅ๐‘–๐‘–

โ€ฒ 2

Page 4: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Top-down Sampling

13

Hyperparameters of the MRGP at the higher scale are considered as the responses of the MRGP at thelower scale.

โ€ข Hierarchy of random fields.โ€ข Non-stationary random field

at the fine scale.โ€ข No limit on the number of

levels/scales.

โ€ข Superscripts โ†’ Level โ€ข Subscripts โ†’ Counter (e.g., for realization)

Level 1 (Coase Scale) Level 2 (Fine Scale)

Needs ๐œท๐œท,๐œฎ๐œฎ,๐Ž๐Žfor its MRGP

Needs ๐œท๐œท,๐œฎ๐œฎ,๐Ž๐Žfor its MRGP

Macroscale MRGP๐’€๐’€๐Ÿ๐Ÿ~๐‘€๐‘€๐‘€๐‘€๐บ๐บ๐บ๐บ ๐œท๐œท๐Ÿ๐Ÿ,๐œฎ๐œฎ1โจ‚๐’“๐’“๐Ÿ๐Ÿ ๏ฟฝ,๏ฟฝ

Realization 1

Realization 2

๐’š๐’š๐Ÿ๐Ÿ๐Ÿ๐Ÿ๐œท๐œท๐Ÿ๐Ÿ,๐œฎ๐œฎ2, ๐’“๐’“๐Ÿ๐Ÿ

๐’š๐’š๐Ÿ๐Ÿ๐Ÿ๐Ÿ๐œท๐œท๐Ÿ๐Ÿ,๐œฎ๐œฎ2, ๐’“๐’“๐Ÿ๐Ÿ

๐œ‡๐œ‡๐‘Œ๐‘Œ = ln๐œ‡๐œ‡๐‘‹๐‘‹

1 + ๐œŽ๐œŽ๐‘‹๐‘‹2๐œ‡๐œ‡๐‘‹๐‘‹2

๐œŽ๐œŽ๐‘Œ๐‘Œ2 = ln 1 +๐œŽ๐œŽ๐‘‹๐‘‹2

๐œ‡๐œ‡๐‘‹๐‘‹2Transformation: required if the underlying distribution is not normal.

โ€ข For ๐‘Œ๐‘Œ normal and ๐‘‹๐‘‹ = exp ๐‘Œ๐‘Œ :

Page 5: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Sensitivity Analysis of Cascaded Effects to the Mesoscale: Moduli

14

โ€ข The average (๏ฟฝฬ…๏ฟฝ๐‘ฃ) is important.โ€ข Spatial variations are not important.

1. Fiber volume fraction: 2. Misalignment angle:โ€ข The average zenith angle (๏ฟฝฬ…๏ฟฝ๐œƒ) is important.โ€ข Spatial variations of ๐œƒ๐œƒ are not important.โ€ข The average and spatial variations of azimuth angle (๐œ‘๐œ‘) are not important.

๏ฟฝฬ…๏ฟฝ๐‘ฃ ๏ฟฝฬ…๏ฟฝ๐œƒ๏ฟฝ๐œ‘๐œ‘

MPaNOTE: Each point inthe figure is averagedover 30 simulations

Sensitivity analyses for dimension reduction: Identify the hyperparameters that affect thehomogenized response of a woven RVE:โ€ข ๐œท๐œท important โ†’Mean values importantโ€ข ๐œฎ๐œฎ or ๐Ž๐Ž important โ†’ Spatial variations important

Page 6: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Microscale & Mesoscale Metamodels

15

3๏ฟฝ๐‘ช๐‘ช =

๐ถ๐ถ1 ๐ถ๐ถ7 ๐ถ๐ถ9๐ถ๐ถ7 ๐ถ๐ถ2 ๐ถ๐ถ8๐ถ๐ถ9 ๐ถ๐ถ8 ๐ถ๐ถ3

0 0 00 0 00 0 0

0 0 00 0 00 0 0

๐ถ๐ถ4 0 00 ๐ถ๐ถ5 00 0 ๐ถ๐ถ6

Stiffness Matrix of UD-RVE

2๏ฟฝ๐‘ช๐‘ช =

๐ถ๐ถ1 ๐ถ๐ถ7 ๐ถ๐ถ11๐ถ๐ถ7 ๐ถ๐ถ2 ๐ถ๐ถ8๐ถ๐ถ11 ๐ถ๐ถ8 ๐ถ๐ถ3

๐ถ๐ถ13 0 0๐ถ๐ถ12 0 0๐ถ๐ถ9 0 0

๐ถ๐ถ13 ๐ถ๐ถ12 ๐ถ๐ถ90 0 00 0 0

๐ถ๐ถ4 0 00 ๐ถ๐ถ5 ๐ถ๐ถ100 ๐ถ๐ถ10 ๐ถ๐ถ6

Stiffness Matrix of the Woven RVE

log(๐‘’๐‘’)

Size of the training dataset

Prediction Error

Micro: Predicting the stiffnessmatrix of a UD- RVE for anyvalues of 3๏ฟฝฬ…๏ฟฝ๐‘ฃ,๐ธ๐ธ๐‘’๐‘’,๐ธ๐ธ๐‘š๐‘š.

โ€ข MRGP training: 30 to 80 samples

โ€ข Validation: 20 samples

Meso: Predicting the stiffness matrix of a woven-RVE for any values of ๐›ผ๐›ผ, 2๏ฟฝฬ…๏ฟฝ๐‘ฃ and 2๏ฟฝฬ…๏ฟฝ๐œƒ.

โ€ข MRGP training: 10 to 60 samples

โ€ข Validation: 20 samples

๐‘’๐‘’ = 1001

20๏ฟฝ๐‘–๐‘–=1

20

1 โˆ’๏ฟฝ๐‘ฆ๐‘ฆ๐‘ฆ๐‘ฆ

2

%

Prediction Error

๐‘’๐‘’

Size of the training dataset

Page 7: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Uncertainty Impacts at the Macroscale

16

Effect on global response: Reaction force(17%) affected by the spatial variations.

Effect on local stress in the mid-sectionโ€ข The coefficient of variation 60

600ร— 100 = 10%.

Page 8: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Comparison of Macroscale Stress Fields

17

MeanStressField

Standard Deviation of Stress Field (30

Simulations)

Ref

๐‘€๐‘€๐บ๐บ๐‘€๐‘€

๐œถ๐œถ AllUncorrelated

AllCorrelated

โ€ข Spatial variations of yarn angle are more important than the other parameters.

โ€ข Correlated spatial variations result in higher stress values.

โ€ข Ref: Reference solution without any spatial variations.

โ€ข The only difference between the cases is spatial variations (i.e., the averages are the same).

Page 9: Nested Multiscale Simulationsuq-materials2019.usacm.org/sites/default/files/20...ย ยท 2019. 2. 13.ย ยท ๐‘ฅ๐‘ฅ. 1. ๐‘ฅ๐‘ฅ. 2. Multi-Response Gaussian Processes (MRGPs) for Quantifying

Remarks on Our Approach

18

โ€ข Non-intrusive: No need to edit the finite element codes

โ€ข Uncertainty Sources: Accounting for multi-response non-stationary uncertain sources

โ€ข Multiscale: Effect of micro and meso uncertainties on macro response

โ€ข Physical insights: MRGP hyperparameters can be directly linked to physical parameters and physical constraints

โ€ข Impact: Uncertain impact will be more critical for nonlinear behavior