Automatic Differentiation at RollsRoyce

1
Automatic Differentiationat Rolls-Royce

• Leigh Lapworth,
• Rolls-Royce PLC, Derby.

2
The HYDRA CFD Solver

• Background
• First version developed by Giles and co-workers
  at OUCL with funding and technical assistance
  from RR
• Developed from outset to be a suite of
  non-linear, linear and adjoint CFD solvers
• Linearised unsteady and adjoint versions were
  hand coded
• Since 2000, has been developed collaboratively by
  RR and a number of university collaborators
  including OUCL.
• Central code in Rolls-Royce corporate CFD
  strategy
• Current Status
• HYDRA used on a wide range of gas turbine
  projects in both the aerospace and non-aerospace
  sectors.
• HYDRA is the basis of a large number of
  industry-university collaborations involving RR.
• The HYDRA collaborative network is one of the
  codes major strengths and the network is
  continuing to grow.

3
HYDRA Framework
Common i/o, parallel, multigrid, visualisation
libraries
Non-Linear Solver
Linearised Unsteady Solver
Steady Adjoint Solver
Steady CFD Unsteady CFD
Steady design sensitivites Loss, Flow rate
etc.
Fixed frequency unsteadiness Flutter, Forced
Response, Tone Noise.
Harmonic Adjoint Solver
Unsteady design sensitivites Work sum, Acoustic
energy, etc.
4
The Role of Automatic Differentiation

• Code maintenance
• Features are added to the non-linear code more
  rapidly than the linear and adjoint versions.
• Inconsistencies between versions affect stability
  and accuracy of the linear and adjoint
  sensitivities.
• Expertise needed to maintain hand-coded versions.
• Automatic Differentiation
• Allows consistent non-linear, linear and adjoint
  versions of the code to be maintained with
  minimal effort.
• More quality assured because scope for
  hand-coding bugs in linear and adjoint versions
  is removed.
• Less validation of new versions needed.
• Generation of linear and adjoint code becomes a
  transient operation controlled by the Makefile
• Only the non-linear code and executables need to
  be retained,
• Less reliance on specialist hand-coding
  expertise.

5
The Influence of Coding Inconsistencies
Inconsistency in flux evaluations leads to larger
errors in adjoint sensitivities
Inconsistency in periodic boundary condition
leads to divergence
6
Makefile Implementation

• AD (Tapenade) automatically run by Makefile when
  linear or adjoint object file needed to build an
  executable.
• Intermediate (AD generated) files are transient
• Sample Makefile

flux.o routines.F CPP -E -C -P routines.F gt
routines.f TPN -forward \ -head
flux \ -output flux \ -vars
q res \ -outvars q res \ routines.f
FC FFLAGS -c flux_d.f /bin/rm -f routines.f
flux_d.f .msg
Giles et. al, Post SAROD Workshop 2005
7
Rolls-Royce Experience with AD

• Acknowledgement
• RR experience has been gathered through research
  programmes with Oxford (Giles et al.), Cambridge
  (Radford) and Cranfield (Forth)
• RR perspective
• Reference hand-coded versions essential to
  demonstrations
• Provided validation and performance database,
• Provided infrastructure so that AD can be applied
  at a subroutine level.
• Application of AD via Makefile essential to
  deployment
• RR doesnt want to maintain an expertise in AD,
  just as it doesnt have experts in compiler
  technology.
• Outlook
• AD is essential to maintaining a large industrial
  code, HYDRA.
• RR will use Tapenade for AD.
• RR will continue to look to collaborators for
  recommendations.

8
Adjoint Design - Bypass OGV design
2D OGV design to reduce static pressure variation
9
Summary

• Automatic Differentiation
• Crucial to long term maintenance of HYDRA,
• AD treated as just another compiler
• Adjoint CFD
• Design
• Several application areas where design variables
  are much greater than objectives and
  constraints.
• Sensitivity analysis
• Often need to ask how close is a design to a
  stability boundary (e.g. stall) or, how
  sensitive is it to small changes (e.g. robust
  design).
• Error analysis
• Either to drive solution adaptation or provide
  confidence levels in solutions.