12/22/2020

Ali M. Sahlodin
Department of Chemical Engineering
AmirKabir University of Technology
1397 S.H

 Stability of ODEs
 Stability of ODE solvers
 Forward (explicit) Euler
 Backward (implicit) Euler

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

2

1
 12/22/2020

 Implicit Euler for DAEs

 Multi-step methods for DAEs
 Backward differentiation formula (BDF)
 Order and stability
 Predictor-corrector method
 Order selection
 Fixed- vs. variable-step BDF

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

3

From Mathworks website

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

4

2
 12/22/2020

Manual car-BDF solver analogy!

Better stability: lower gear/BDF order

Faster driving/integration: Higher gear/BDF order (still accurate with larger

steps)
Both start from the 1st gear/order (why?)

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

5

 Importance of a good initial guess

Cycling solution Solution divergence

Courtesy of Benoit Chachuat

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

6

3
 12/22/2020

 Solver failure

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

7

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

8

4
 12/22/2020

Equations K. E. Brenan, S. L. Campbell, and L. R. Petzold 1995

Numerical Solution of Initial-Value Problems in Differential-Algebraic
 Corrector equation

 Modified Newton’s method

 Iteration matrix

Iteration matrix evaluated at predicted y and y’

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.
9

 Iteration matrix: dominant cost of integration

 Partial differentiation
 Matrix inversion
 Simplify matrix inversion: LU factorization

 How to reduce computational cost?

 Do not update G at every step!
 May converge in more iterations
 But, each iteration is less expensive

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

10

5
 12/22/2020

 Symbolic differentiation
 Exact derivatives
 Difficult/impossible to obtain

 Finite differences
 Approximate derivatives
 Easy to implement

 Automatic differentiation
 Exact derivatives
 Easy to implement and use

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

11

 Break down the function into its constituting

elements

https://en.wikipedia.org/wiki/Automatic_
differentiation

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

12

6
 12/22/2020

Calculate both the function and its derivative at the same time

https://en.wikipedia.org/wiki/Automatic_
differentiation

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

13

 Operator overloading: define new data type that

carries both variable and its derivative
 Evaluate derivatives along with function evaluations
 less efficient
 Easy to implement (C++, object-oriented and template)

 Source code transformation: generate derivative

code from original program
 More efficient
 More difficult to implement

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

14

7
 12/22/2020

http://www.autodiff.org/

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

15

 Set your solver to use user-supplied

derivatives (instead of finite difference)
 Have the automatic differentiation package
take the function derivatives
 Supply the derivatives to the solver

Advanced simulation packages have built-in automatic differentiation

capabilities
User provides the original model only!

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

16

8
 12/22/2020

y  y,   diag (107 , 107 ,..., 107 ), n  5000

Without user-supplied Jacobian With user-supplied Jacobian

94 successful steps
0 failed attempts 94 successful steps
5190 function evaluations 0 failed attempts
1 partial derivatives 190 function evaluations
18 LU decompositions 1 partial derivatives
188 solutions of linear systems 18 LU decompositions
CPU time = 369.3594 188 solutions of linear systems
Elapsed time is 100.947378 CPU time = 81.8750
seconds. Elapsed time is 28.116323 seconds.

•Finite differencing requires a Large number of function evaluations

•User-supplied Jacobian saves a lot of time for large systems! (how?)

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

17

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

18

9
 12/22/2020

From “Nonlinear and Dynamic Optimization”, Benoit Chachuat

 Octave (daspk, dassl, ode15s, ode15i)
 Modelica (OpenModelica)
 MATLAB (ODE15i, ODE15s)

Copyright © Ali M. Sahlodin, Dept. of Chemical Engineering, AmirKabir Univ. of Tech.

19

10