Solving PDE related problems using deep-learning

Post on 16-Jan-2022

5 views 0 download

Transcript of Solving PDE related problems using deep-learning

Solving PDE related

problems using

deep-learning

Adar KahanaTel Aviv University

Under the supervision of

Eli Turkel (TAU) and

Dan Givoli (Technion)

Shai Dekel (TAU)

Waves seminar,

UC Merced

February 4th , 2021

1

Agenda

Motivation

Data driven problems

Obstacle identification and deep-learning

Dealing with CFL instability using deep-learning

2

Underwater acoustics

Sonar imaging

3

The wave problem

4

แˆท๐‘ข ิฆ๐‘ฅ, ๐‘ก = ๐‘‘๐‘–๐‘ฃ ๐‘ ิฆ๐‘ฅ 2๐›ป๐‘ข ิฆ๐‘ฅ, ๐‘ก , ิฆ๐‘ฅ โˆˆ ฮฉ, t โˆˆ (0, ๐‘‡]

๐‘ข ิฆ๐‘ฅ, 0 = ๐‘ข0 ิฆ๐‘ฅ , ิฆ๐‘ฅ โˆˆ ฮฉ

แˆถ๐‘ข ิฆ๐‘ฅ, 0 = แˆถ๐‘ข0 ิฆ๐‘ฅ , ิฆ๐‘ฅ โˆˆ ฮฉ

๐‘ข ิฆ๐‘ฅ, ๐‘ก = ๐‘“ ิฆ๐‘ฅ, ๐‘ก , ิฆ๐‘ฅ โˆˆ ๐œ•ฮฉ1, t โˆˆ 0, ๐‘‡

๐›ป๐‘ข ิฆ๐‘ฅ, ๐‘ก = ๐‘” ิฆ๐‘ฅ, ๐‘ก , ิฆ๐‘ฅ โˆˆ ๐œ•ฮฉ2, t โˆˆ [0, ๐‘‡]

๐œ•ฮฉ1 โˆช ๐œ•ฮฉ2 = ๐œ•ฮฉ

where แˆถ๐‘ข0 ิฆ๐‘ฅ = 0 and ๐‘“ ิฆ๐‘ฅ, ๐‘ก = ๐‘” ิฆ๐‘ฅ, ๐‘ก = 0

Ill-posed problems

In an experiment, we store the pressure at a small

number of sensors for all time steps

We wish to find the properties of the source or

obstacle from the data stored at these sensors

where the number of sensors << mesh

This is an inverse problem which is highly ill-posed

Hence, one cannot usually reconstruct the initial

conditions perfectly

Can we solve these types of ill-posed problems

with learning?

5

Partial information โ€œRecordingโ€ the solution at a small set of

sensors placed in the domain ิฆ๐‘ฅs๐‘› ๐‘›=1

๐พโˆˆ ฮฉ

Data โ€“

๐‘ข ิฆ๐‘ฅs1 , t

๐‘ข ิฆ๐‘ฅs2 , t

โ‹ฎ๐‘ข ิฆ๐‘ฅsn , t

+ ๐’ฉ ๐œ‡, ๐œŽ2

The ill-posedness raises sensitivity to noise at

the sensors

6

Agenda

Motivation

Data driven problems

Obstacle identification and deep-learning

Dealing with CFL instability using deep-learning

7

Data driven problems

Supervised learning

Input data

Output labels

Training

Prediction (testing)

Drawback - sensitivity

8

Deep-learning

Training โ€œweightsโ€ to learn connections in the data

Hidden multi-dimensional embeddings

Convolutions, Fully connected

โ€œDeepโ€ and non-linear

Loss

9

1 0 10 1 01 0 1

โ†’

๐‘ค1 ๐‘ค2 ๐‘ค3

๐‘ค4 ๐‘ค5 ๐‘ค6

๐‘ค7 ๐‘ค8 ๐‘ค9

Physically-informed NN

Input: set of points from the initial and boundary

conditions

Output: solution in the domain

Loss: the problem

10

M. Raissi, P. Perdikaris, G.E. Karniadakis, Journal of computational Physics, 2018

Results

11

Agenda

Motivation

Data driven problems

Obstacle identification and deep-learning

Dealing with CFL instability using deep-learning

12

Problem definitionGiven the position of the source/s and data at a few

sensors but many time slices find the location, size and

shape of the unknown scatterers

Input: Sensors recordings (๐‘๐‘ ๐‘Ž๐‘š๐‘๐‘™๐‘’๐‘  ร— ๐‘๐‘ก๐‘ ๐‘ก๐‘’๐‘๐‘  ร— ๐‘๐‘ ๐‘’๐‘›๐‘ ๐‘œ๐‘Ÿ๐‘ )

Output: Obstacle?

13

Prior work

Location ิฆ๐‘ฅ

Shape and size โ€“

Circles: Radius

Rectangles: Height and Width

Complex shapes: need to be parametrized

โ€œSoftโ€ obstacles โ€“

Semi-penetrable

Multiphysics

14

Labels solution - segments

15

Labels are ๐‘š ร— ๐‘›binary matrices

Predictions will be

๐‘š ร— ๐‘› probability

matrices

Loss: NLL

Spatio-temporal architecture

16

Loss diagram

17

Probability

map

Physically informed loss

Using the segmentation network and output เทจ๐‘‚

Define a loss component based on:

Solve: ๐‘ข๐‘ก๐‘ก = 1 โˆ’ เทจ๐‘‚ ๐‘ฅ ๐‘2(๐‘ฅ) ฮ”๐‘ข

Get sensor data: ๐‘ข๐‘˜ ๐‘ฅ๐‘ ๐‘– ๐‘–=1

#๐‘ ๐‘’๐‘›๐‘ ๐‘œ๐‘Ÿ๐‘ for each sample

Calculate MSE between ground truth

๐‘ข๐‘˜ ๐‘ฅ๐‘ ๐‘– ๐‘–=1

#๐‘ ๐‘’๐‘›๐‘ ๐‘œ๐‘Ÿ๐‘ and the prediction as component ๐‘™2

Define the loss function for our network as:

๐›ผ โ‹… ๐‘™1 + 1 โˆ’ ๐›ผ โ‹… ๐‘™2such that ๐‘™1 is the NLL loss described earlier

18

Numerical experiments Dirichlet BC

Compact Gaussian initial condition

Arbitrary polygonal obstacles โ€“

Generate number of edges

Generate edge length and angle

Generate location (๐‘ฅ0, ๐‘ฆ0, ๐‘ง0)

โ– Enormous samples space

Generated only 25,000 samples

19

Probability images

20

Neural network โ€“ results Intersection over union: 0 โ‰ค ๐ผ๐‘‚๐‘ˆ ๐ด, ๐ต =

๐ดโˆฉ๐ต

๐ดโˆช๐ตโ‰ค 1

Up to 66% IOU score

21

A.K., E. Turkel, D. Givoli, S. Dekel, Journal of computational Physics, 2020

Agenda

Motivation

Data driven problems

Obstacle identification and deep-learning

Dealing with CFL instability using deep-learning

22

Explicit schemes and CFL One-dimensional wave equation

CFL condition for stability: ๐›ผ =๐‘ฮ”๐‘ก

ฮ”๐‘ฅโ‰ค 1

FDCD: ๐‘ข๐‘–๐‘›+1 = 2๐‘ข๐‘–

๐‘› โˆ’ ๐‘ข๐‘–๐‘›โˆ’1 + ๐›ผ2 ๐‘ข๐‘–+1

๐‘› โˆ’ 2๐‘ข๐‘–๐‘› + ๐‘ข๐‘–โˆ’1

๐‘›

23

Network architecture

Input: ๐‘ข ๐‘›โˆ’1 ๐‘š, ๐‘ข๐‘›๐‘š

Output: ๐‘ข ๐‘›+1 ๐‘š

Spatio-temporal architecture

Non-linear activation (PReLU)

Loss: MSE between ๐‘ข ๐‘›+1 ๐‘š and เทœ๐‘ข ๐‘›+1 ๐‘š

24

Network diagram

25

Physics informed loss

Use ๐‘ข ๐‘›โˆ’1 ๐‘š, ๐‘ข๐‘›๐‘š to predict เทœ๐‘ข ๐‘›+1 ๐‘š

Inside the loss:

Use ๐‘ข ๐‘›+1 ๐‘š, ๐‘ข ๐‘›+1 ๐‘š+1 to calculate ๐‘ข ๐‘›+1 ๐‘š+๐‘—

Use เทœ๐‘ข ๐‘›+1 ๐‘š, ๐‘ข ๐‘›+1 ๐‘š+1 to predict เทœ๐‘ข ๐‘›+1 ๐‘š+๐‘—

Calculate the MSE between ๐‘ข ๐‘›+1 ๐‘š+๐‘— and เทœ๐‘ข ๐‘›+1 ๐‘š+๐‘—

Network loss is the linear combination of the two

MSE losses

26

Numerical experiments Dirichlet BC

Data:

Linear combinations with random coefficients

created from the basis sin ๐œ‹๐‘˜๐‘ฅ ๐‘˜=120

1250 different initial condition and 397 time-steps for

each one, total of 496,250 samples

Samples created with CFL = 0.875 and only each

10th sample was taken to get CFL = 8.75

27

Results

28

Results29

O. Ovadia, A. K, E. Turkel, S. Dekel, Journal of computational physics, submitted

Summary and future work

Obstacle location and identification

Investigating source location

High measurement noise

Stability

Extending to 2,3 dimensions

Dispersion relation problem โ€“ optimized kernels

Experimental data

30

Thanks!

31