Automatic Differentiation and Its Role in Simulation-Based Optimization

Outline

Group Members

AD in a Nutshell

“Black Box” Methods

Multidisciplinary Design Optimization of Airfoils

Automatic Tuning of Sea Ice Model Parameters

IABP buoy data

Ice drift

Contours of ice drift speed

Ice thickness distribution

AD of Fluent

Slide 13

Fluent Sample Problem

Fluent Sample Problem (cont)

FLUENT.AD - first results

FLUENT.AD - first results (con‘t)

Issues with Black Box Differentiation

Wisconsin Sea Ice Model

Influence of Perturbations on Film Problem (Aachen)

Influence of compilers (single precision)

Difficulties during Fluent Preprocessing

Overflows in Fluent.AD

“Gray Box” Methods

SensPVODE: Objective

Possible Approaches

PVODE as ODE + sensitivity solver

SensPVODE: Test Problem

SensPVODE: Number of Timesteps

SensPVODE: Time/Timestep

Differentiated PETSc Linear Solver (SLES)

AD, CD – More Accurate Than DD

All-at-once (SAND) methods

LNKS: parameter identification model problem

Automating AD (User’s Perspective)

AD/PETSc Automation

Tools: ADIC 2.0

R&D Activities

Addressing Limitations in Black Box AD

Hybrid AD/FD Method (switching)

Hybrid AD/FD Method (Turner-Walker)

Hybrid AD/FD Method (Combination)

Conclusions