09:00-9:20

Welcome Prof. Kyriakos Giannakoglou, NTUA, Dr. Jens-Dominik Mueller, QMUL

• Organisation of the workshop, schedule, programme (KG)
• Overview of the About Flow project (JDM): Motivation and rationale; aims, objectives, structure; training provided by the project; deliverables

WS Part 1: Introduction to Adjoint-Based Optimisation

09:20-10:30

1: Numerical Optimisation (JDM)

• Overview: what is it, what can it do, what are the benefits, where are the difficulties Examples.  Stochastic vs. gradient based optimisation. Advantages of gradient-based optimisation in CFD, issues with industrial application.
• Overview of popular gradient-based optimisation algorithms. Gradient computation, assembling derivatives, chain rule. Parametrisation approaches. Workflows: the design chain

2: Introduction to adjoint methods (JDM)

• Discrete vs continuous adjoints, gradient computation
• physical meaning of adjoint fields
• examples of sensitivity fields

14:00-15:30

3: Numerical optimisation in industry I

• General industrial applications (Dr. Mustafa Megahed (ESI-Group), Dr. Guillaume Pierrot (ESI-Group)): Currently practiced optimisation methods. Typical workflow examples, arising bottlenecks. What is currently possible, where are the challenges, what is needed?
• Adjoint-based optimisation for car development: Methods and Applications (Dr. Carsten Othmer, Volkswagen AG). Topology and shape optimisation for ducted flows, optimisation of external aerodynamics: Sensitivity maps for shape optimisation and flow control, challenging applications for unsteady flow

4: Numerical optimisation in industry II

• Applications in turbomachinery (Dr. Marius Swoboda (Rolls Royce Deutschland)). Currently practiced optimisation methodsi in turbomachinery design. Typical workflow examples, arising bottlenecks. What is currently possible, where are the challenges, what is needed in industrial optimisation?

09:00-10:30

5: AD Tools (Dr. Laurent Hascoet, INRIA; Prof. Uwe Naumann, RWTH Aachen)

• Discrete adjoint versions of numerical simulation code can be generated by a semantic code transformation technique known  as Algorithmic Differentiation (AD). Tools for AD are distinguished by their fundamental approach to the modification of the code's semantics. Source-to-source transformation tools (e.g. Tapenade developed by INRIA) rewrite a given (C or Fortran in the case of Tapenade) code using standard compiler techniques. Overloading tools (e.g. dco developed by STCE) use function and operator overloading techniques (in C++ and Fortran in the case of dco) to achieve a similar effect. Both approaches yield discrete adjoint numerical simulation code of first and potentially higher order.
• This introductory lecture gives a brief overview of AD techniques and of the corresponding tools (Tapenade and dco). A motivating case study is followed by "Hello World" examples illustrating the source-to-source and overloading approaches to AD, respectively. Brief "How to get started with Tapenade/dco" guides form the basis for more in-depth discussions of AD (tools) in a dedicated upcoming workshop organised by the About Flow project in 2014.

11:00-12:30

6: Adjoint CFD codes (JDM)

• Overview of application of AD to CFD codes, possible approaches and their advantages and disadvantages. Discussion of challenges that are typically encountered and how to overcome them. Examples of successfully differentiated CFD codes and the resulting performance and robustness.

AboutFlow Steering Committee Meeting

7: Parametrisation for shape optimisation (JDM, SX(QMUL))

• Overview of methods: lattice based, free-form deformation, radial-basis function
• Node-based parametrisations, regularisation, preconditioning.
• CAD-based parametrisations

11:00-12:30

8: Development and Application of Optimization Methods beyond Adjoint Methods – The EASY Platform ( Kyriakos Giannakoglou, NTUA)

Excursion - Island Tour

WS Part 2: Continuous Adjoint

09:00-10:30

9: Introduction to Continuous Adjoint Methods (Kyriakos Giannakoglou, NTUA)

• The talk will give a detailed introduction to the continuous adjoint method, based on some quasi-one-dimensional flow examples. Both compressible and incompressible flows will be covered, by focusing on the particular features of each one of them. The field adjoint equations, their boundary conditions as well as expressions for the sensitivity derivatives will be derived

11:00-11:30

A Physicist’s Perspective on the Adjoint Method (Carsten Othmer, VW)

• An attempt for an illustrative explanation of the adjoint method
• The “physics“ behind a simple adjoint problem: heat conduction
• Mathematical justification

10: Continuous Adjoint Methods for Turbulent Flows with Industrial Applications (Dr. Evangelos Papoutsis-Kiachagias, NTUA)

• The talk will cover the process of deriving the adjoint equations to both low- and high-Re number turbulence models using the continuous adjoint method. The differences with the commonly made “frozen turbulence” assumption and the gain from the exact mathematical derivation will be highlighted.  Automotive and turbomachinery applications will be presented.

14:00-15:30

11: Development of a Multi-Function Continuous Adjoint Solver Platform in OPENFOAM  (Dr. Eugene de Villiers, Engys)

• The talk will cover several aspects of the adjoint optimisation developed in partnership with VW – topology optimisation, multi-objective optimisation and DES aerodynamic surface sensitivity – but will focus on the platform and software implementation.
  • Specific items that will be covered:
  • General extensible framework for running different types of primal/adjoint solver combinations
  • Multi-objective framework
  • Automation
  • Manufacturability and other simple constraints
• The aspects of OPENFOAM that are advantageous for such an implementation will be highlighted in the context of each component (central database, high level abstractions, flexible library of complimentary functionality)
• Current developments and future directions of the platform

12: The Continuous Adjoint Method on CPUs & GPUs for Multi-level & Multi-objective Optimization (Dr. Varvara Asouti, NTUA)

• The first part of this talk will expand previous presentations on the continuous adjoint method, based on CFD codes other than OpenFOAM, particularly those which are enabled to efficiently run on GPUs. Codes running on GPUs offer the indisputable advantage of very high speed-up compared to their counterparts running on CPUs. This gain is important if optimization loops are to be performed. The talk includes the use of preconditioning techniques and poses the dilemma of choosing between the use of the preconditioned adjoint method or the adjoint to the preconditioned primal equations.
• The second part of the talk is about smart ways to exploit the availability of an adjoint solver by hybridizing it with other search methods, leading to very efficient multi-level optimization loops. The talk focuses on the use of continuous adjoint to solve multi-objective optimization problems and presents a way to reduce the number of calls to the adjoint solver. In fact, it demonstrates how a single call (rather than as many as the number of objectives) to the adjoint solver might be enough.

09:00-10:30

13: On the Use of Adjoint Methods for the Computation of Higher-Order Sensitivity Derivatives & Robust Optimization (Prof. Kyriakos Giannakoglou, NTUA)

• The application of continuous adjoint methods for the computation of the Hessian of the objective function F is presented. The Hessian is needed to support the use of the (exact) Newton method. A comparison with the corresponding discrete adjoint methods follows. Other efficient alternatives, such as the exactly-initialized quasi-Newton method or truncated Newton methods are also presented.
• In the second part of the talk, the implementation of the adjoint formulation for the solution of robust design problems in aerodynamic shape optimization, with affordable CPU cost, is presented. In robust design, the minimization of both the mean value and standard deviation of F takes into account the so-called environmental variables. According to the second-order, second-moment approach, first and second-order derivatives of F with respect to the environmental variables must be computed. To perform a gradient-based optimization, these quantities must be differentiated again, with respect to the shape controlling variables. Thus, methods for computing up to third-order mixed sensitivities of F are needed and these will be exposed in this part of the talk.

11:00-12:30

14: Continuous Adjoint Methods for Topology Optimization, Flow Control & Unsteady Flows (Dr. Evangelos Papoutsis, NTUA)

• The talk will cover the process of deriving the continuous adjoint formulation for topology and flow control optimization problems, along with some indicative applications. The challenges faced when dealing with the adjoint to unsteady flows will be presented along with the possible remedies.