35
Systematic stochastic modeling of atmospheric variability Christian Franzke Meteorological Institute Center for Earth System Research and Sustainability University of Hamburg Daniel Peavoy and Gareth Roberts (Warwick) Daan Crommelin (CWI) and Andy Majda (NYU)

Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

Systematic stochastic modeling

of atmospheric variability

Christian FranzkeMeteorological Institute

Center for Earth System Research and Sustainability 

University of Hamburg

Daniel Peavoy and Gareth Roberts (Warwick)

Daan Crommelin (CWI) and Andy Majda (NYU)

Page 2: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  2

Outline

1) Motivation

2) Stochastic Mode Reduction

3) Physical Constraints

4) Bayesian Parameter Estimation

5) Results

6) Summary

Page 3: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  3

Scales

Page 4: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  4

Long­Range Forecasts

Page 5: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  5

Reduced Stochastic Climate Models

 → Computationally much cheaper

 → Capture essential dynamics

­ Improved extended range forecasting­ Large ensemble forecasting­ Long­term climate studies (e.g. paleo­climate)­ Long control simulations to estimate extremes­ Extreme Event Prediction

Page 6: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  6

Reduced Order Models

Slow Climate Modes

Fast Weather Modes

Reduced Model

Assumption: Time scale separation

Page 7: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  7

Reduced Order Models

Assumption: Time scale separation

Page 8: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  8

Stochastic Modeling

Page 9: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  9

Outline

1) Motivation

2) Stochastic Mode Reduction

3) Physical Constraints

4) Bayesian Parameter Estimation

5) Results

6) Summary

Page 10: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  10

Stochastic Mode Reduction

           Equations of motions         Energy conservation 

Decompose u into (x,y)

x: slow compomenty: fast component

dudt

=F+Lu+ I (u ,u) u∗I (u ,u)=0

Page 11: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  11

Stochastic Mode Reduction

Fast nonlinear interactions: I(y,y)

e.g. Ohrnstein­Uhlenbeck Process → Continous time version of AR(1)

Majda et al.1999, 2001, 2008; Franzke et al. 2005; Franzke and Majda 2006

Page 12: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  12

Stochastic Mode Reduction

           Equations of motions         Energy conservation 

dudt

=F+Lu+ I (u ,u) u∗I (u ,u)=0

dx=(~F+

~L x+~I (x , x)+M (x , x , x))dt+σA dW A+σM (x)dW M

Reduced Model:

Mode Reduction

Page 13: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  13

Stochastic Mode Reduction

Page 14: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  14

Stochastic Mode Reduction

Solve for y:

Page 15: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  15

Stochastic Mode Reduction

For

Page 16: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  16

Stochastic Mode Reduction

Plug into equation for x:

Page 17: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  17

Stochastic Mode Reduction

Plug into equation for x:

CAM Noise

Page 18: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  18

Stochastic Mode Reduction

Plug into equation for x:

CAM Noise

Cubic Term

Page 19: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  22

Outline

1) Motivation

2) Stochastic Mode Reduction

3) Physical Constraints

4) Bayesian Parameter Estimation

5) Results

6) Summary

Page 20: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  23

Constraints on Stochastic Climate Models

Page 21: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  24

Constraints on Stochastic Climate Models

Only considering cubic terms

Majda et al. 2009; Peavoy et al. 2015

Page 22: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  25

Constraints on Stochastic Climate Models

Majda et al. 2009; Peavoy et al. 2015

Stability: 

Quadratic form Q is negative­definite

Allows the system to be linearly unstable

Page 23: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  26

Peavoy et al. 2015

Physical Constraints

Without constraint about40% of parameterestimates lead to unstablesolutions

Page 24: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  27

Outline

1) Motivation

2) Stochastic Mode Reduction

3) Physical Constraints

4) Bayesian Parameter Estimation

5) Results

6) Summary

Page 25: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  28

Bayesian Parameter Estimation Procedure

Imputing of Data

Discretisation: Euler­Maruyama scheme

Modified Linear Bridge

Likelihood based parameter estimation (MCMC)

Peavoy et al. 2015

Page 26: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  29

Peavoy et al. 2015

How to sample negative­definite matrices?

Wishart Distribution

Truncated Normal Algorithm:A n×n matrix is negative definite if and only if all k≤n 

leading principal minors obey |Mk|(­1)k > 0. The k­th principal minor is the determinant of the upperleft k×k sub­matrix.

Diagonal ElementsOff­Diagonal Elements

Page 27: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  30

Peavoy et al. 2015

Modelling Memory Effects via Latent Variables

Red Noise

Page 28: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  31

Outline

1) Motivation

2) Stochastic Mode Reduction

3) Physical Constraints

4) Bayesian Parameter Estimation

5) Results

6) Summary

Page 29: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  32

Triad Model Example

ModelReduction

Peavoy et al. 2015

Page 30: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  33

Triad Model Example

Peavoy et al. 2015

Page 31: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  34

Test: Chaotic Lorenz Model

ε = 0.1 ε = 0.01

Reduced order model:

Peavoy et al. 2015

Page 32: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  35

Peavoy et al. 2015

Flow over topography on a ß­plane

Page 33: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  36

Arctic Oscillation Index 

Autocorrelation Function

From NCAR­NCEP reanalysis datacovering period 1948­2010

Page 34: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  37

● North Atlantic Jet Stream has three persistent states

● Persistent states exhibit variability on interannual and decadal time scales

● Propensity of extreme wind speeds depend on persistent states

Page 35: Systematic stochastic modeling of atmospheric …38 Summary Normal form for reduced stochastic climate models predict a cubic nonlinear drift and a correlated additive and multiplicative

  38

Summary● Normal form for reduced stochastic climate models 

predict a cubic nonlinear drift and a correlated additive and multiplicative CAM noise.

● Bayesian Framework for Physics Constrained Parameter Estimation 

● Reduced stochastic climate models perform wellReferences:● Majda, Franzke and Crommelin, 2009: Normal forms for reduced stochastic 

climate  models. Proc. Natl. Acad. Sci. USA, 106, 3649­3653.● Peavoy, Franzke and Roberts, 2015: Physics constrained parameter estimation of 

stochastic differential equations. Comp. Stat. Data Ana., 83, 182­199.● Franzke, C., T. O'Kane, J. Berner, P. Williams and V. Lucarini, 2015: Stochastic 

Climate Theory and Modelling. WIREs Climate Change, 6, 63­78.● Gottwald, G., D. Crommelin and C. Franzke, 2016: Stochastic Climate Theory. To 

appear in Nonlinear and Stochastic Climate Dynamics, Cambridge University Press.