Using Machine Learning for Epistemic Uncertainty Quantification in Combustion and Turbulence Modeling 1.

1

Using Machine Learning for Epistemic Uncertainty

Quantification in Combustion and Turbulence Modeling

2

Epistemic UQ

• Use machine learning to learn the error between the low fidelity model and the high fidelity model– Want to use it as a correction and an estimate of error

• Working on two aspects -- Approximate the real source term (in progress equation)

given a RANS+FPVA solution– Approximate the real Reynolds stress anisotropy given an

eddy-viscosity based RANS solution • Preliminary work

– We will show a way it could be done, not how it should be done

3

Basic Idea

• We can compare low fidelity results to high fidelity results and learn an error model– Model answers: “What is the true value given the low-

fidelity result”• If the error model is stochastic (and correct), draws

from that model give us estimates of uncertainty. • To make model fitting tractable we decouple the

problem– Model of local uncertainty based on flow-features– Model of coupling of uncertainty on a macro scale

4

Local Model

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-3

-2

-1

0

1

2

3

Fake Low Fidelity Feature

Fak

e H

igh

Fid

elity

Fea

ture

5

Model Generation Outline

• Get a training set which consists of low-fidelity solutions alongside the high-fidelity results

• Choose a set of features in high-fidelity to be learned ( y )

• Choose a set of features in low-fidelity which are good representations of the error ( x )

• Learn a model for the true output given the input flow features

6

Example

• In the RANS/DNS case, we are interested in the RANS turbulence model errors

• Input of the model is RANS location of the barycentric map, the marker, wall distance, and (5 dimensional)

• Output of the model is DNS location in the barycentric map (2 dimensional)

7

Local Model

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-3

-2

-1

0

1

2

3

Fake Low Fidelity Feature

Fak

e H

igh

Fid

elity

Fea

ture

8

Sinker

• For a test location, each point in the training set is given a weight set by a kernel function

• Then, using the true result at the training points and the weights, compute a probability distribution over the true result

•

9

Example Problem

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-30

-20

-10

0

10

20

30

10

30 Samples

-1.5 -1 -0.5 0 0.5 1 1.5 2-4

-3

-2

-1

0

1

2

3

4

11

100 Samples

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-4

-3

-2

-1

0

1

2

3

12

300 Samples

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-4

-3

-2

-1

0

1

2

3

13

1000 Samples

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-4

-3

-2

-1

0

1

2

3

14

10000 Samples

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2-4

-3

-2

-1

0

1

2

3

4

15

Combustion Modeling

• DNS finite rate chemistry dataset as high fidelity model, RANS flamelet model is low fidelity model

• Input flow features are the flamelet table variables (mixture fraction, mixture fraction variance, progress variable)

• Output flow variable is source term in progress-variable equation

• Use a GP as the spatial fit

16

‘Truth’ Model

Dataset used : Snapshots of temporal mixing layer data from Amirreza

17

Trajectory Random Draws

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

x 10-3

0

500

1000

1500

2000

2500

3000

Location

Sour

ceFPVA Table

18

Initial condition

19

Results of ML scheme

20

Application to EUQ of RANS

21

Input Data

• Add in marker, normalized wall distance, and p/ε as additional flow features, and use Sinker

22

Model Output

23

• Not perfect, but way better

24

Generating Errorbars

• Each point also has a variance associated with it (which is an ellipse for now)

• We can use these uncertainties to generate error bars on macroscopic quantities

• Draw two Gaussian random variables, and tweak the barycentric coordinate by that many standard deviations in x and y

• If the point goes off the triangle, project it back onto the triangle

• Gives us a family of new turbulence models

25

Random Draws

26

Random Draws

27

Conclusions

• Promising early results• Basic idea:

Learn `mean and variance’ of error distribution of modeling terms in the space of FEATURES

• There is a lot of work to be done– Feature selection– Better uncertainty modeling (non-Gaussian)– Kernel selection

• Need to develop a progressive / logical test suite to evaluate the quality of a model