Skip to main content

Industrial optimal design using adjoint CFD

Periodic Reporting for period 2 - IODA (Industrial optimal design using adjoint CFD)

Reporting period: 2017-01-01 to 2018-12-31

Adjoint-based methods are the most interesting approach in numerical optimisation using Computational Fluid Dynamics (CFD). To enable its routine industrial application will be an essential ingredient in maintaining the competitive edge of European manufacturers and to support the EU Knowledge Economy.
Parametrisation is at the core of shape optimisation, but bottlenecks primarily arise as current parametrisations do not provide the shape derivatives needed for adjoint optimisation. IODA addressed two bottlenecks: a) efficient but flexible and automatic parametrisation of arbitrary shapes, and b) imposition of design constraints.
IODA’s progress with parametrisation enables to apply these methods to a much wider range of problems with much faster turnaround times, and ultimately reduced 'time to market' of new products.
A suite of industrial benchmark testcases was defined, where possible underpinned by experimental results. These cases are open to the research community and benchmark exercises will be held at regular intervals.
CAD-free parametrisation methods were advanced with improvements in robustness of surface and volume mesh smoothing methods. Implicit surface smoothing methods were demonstrated on automotive cases. Adaptive explicit/implicit surface smoothing was developed for arbitrary surface topologies. Efficiency improvements were developed for RBF volume deformation methods.
Development of transition tools from CAD-free to CAD-based proved challenging, but working models for 2D could be produced, and first steps were made toward novel approaches for automated topology recognition in 3D.
Substantial work was performed on producing gradient-enabled CAD models that can be integrated into the design loop, hence accepting an initial CAD description and producing the optimal shape in CAD. A range of complementary approaches were developed, a) finite-differences on commercial CAD kernels, b) automatic-differentiation of the open-source CAD kernel OpenCascade, c) differentiation of in-house CAD tools. All of these methods achieved good maturity and were able to handle the benchmark cases.
Gradient-enabled approaches to impose constraints were developed suitable for the various CAD approaches and demonstrated on duct and turbomachinery blade cases. Methods to extend the parametrisation tools to include uncertainty were implemented, including a novel multi-level/multi-fidelity Monte-Carlo approach that is suitable for large numbers of uncertainties.
All the developments were tested on at least one of the benchmark cases. Publications that assess the performance and capabilities of the various approaches are in preparation. All partners have made substantial advances with their in-house capability on adjoint design optimisation. Three major routes of further exploitation are envisaged: a) direct exploitation in the workflows of the OEM partners, b) exploitation through better capability of the software/consultancy partners and c) exploitation through public and industrial research grants by the academic partners.
Good improvements to robustness and adaptive scaling of surface smoothing methods was achieved for CAD-free surface-based methods. Work on RBF volume morphing achieved substantial improvements in computational efficiency. These methods can now be used in routine workflows for design exploration. However, progress with ‘return-to-CAD’ is more limited, this remains a challenging problem. Accepting a loss of generality enables a templating approach which was very successfully demonstrated on the DrivAer mirror case.
Work on CAD-based parametrisation in WP4 was a resounding success with major advances with all approaches that were investigated.
The innovative `implicit’ parametrisation approach NSPCC of QMUL has been applied to a wide range of cases and enhanced with tools for adaptive design space refinement. This approach achieved best results when applied to the S-Bend testcase.
The finite-difference gradient approach of QUB has been very successfully applied to a range of complex industrial cases including constraint handling through generic CAD distance evaluation. Further advances have been demonstrated in adaptive refinement of the CAD feature tree to enrich the design space.
A significant success was the complete differentiation of the open-source CAD kernel Open Cascade Technologies. Both `explicit’ parametrisations, classic engineering approaches that build up the feature tree from points, lines and surfaces, as well as `implicit’ parametrisations that work with the control nets of the NURBS patches forming the Boundary Representation (BRep) have been demonstrated.
Gradient-enabled constraint methodologies were enabled. Most prominent were CAD-based approaches using differentiated distance/collision functionality of the CAD systems, as well as the definition auxiliary constraint functions. The approaches were successfully demonstrated on turbomachinery blades.
IODA also progressed with uncertainty quantification using established polynomial chaos methods, as well as innovative multi-level/multi-fidelity Monte Carlo methods using inexpensive adjoint low-fidelity models. Both approaches are now ready to be exploited in robust engineering design.

All partners have made substantial advances with their in-house capability on adjoint design optimisation. Major routes to impact are:
1. OEM partners (VW, RRD) will use the project developments in their design workflows to develop improved products and hence increase their competitivity in the global maketplace, resulting in employment creation in the EU knowledge economy. The products that can now be designed with much lower environmental impact, in particular CO2 emissions, will make the EU and the world a better place to live.
2. Partners who develop software (Engys, ESI, OCCT) and research institutes (VKI) have significantly increased their capability. The availability of software and services will help smaller and SME companies access this technology, as well as provide employment at the partners. This will have a positive impact on the EU knowledge economy, as well as help to deliver the ambitious environmental targets of the EU.
3. Academic partners (QMUL, NTUA, UPB) have improved their competence and capability, putting them at the forefront of research in the field, and are using this in funding bids.
4. The strong dissemination effort with over 50 open-access papers, as well as the organisation of 3 dedicated minisymposia at leading conferences, makes the developments very accessible to the wider R&D community, supporting strong progress in this field.
5. The very comprehensive training programme has developed a class of 15 researchers with skills to lead in this field.
6. A substantial outreach effort produced a wide range of activities, has delivered a strong contribution to getting school children interested in STEM subjects.
Leading edge radius AD grid sensitivities
Constrained Optimization Queen Mary University of London Work Package 5
Robust mesh optimization using optimized smoothing enables exploring richer design space
Application of CAD-free morphing framework on an external aerodynamic case with advanced filtering
Shape Optimization Work Packages 2 and 3
CAD-parametrisation of a turbo-machinery U-Bend cooling channel
Sensitivity of the leading edge radius of a turbomachinery blade generated by autom. differentiation
Cold-to-Hot CAD deformation of an axial fan blade Von Karman Institute Work Package 4
Maximum von Mises stress sensitivity map of an axial fan blade Von Karman Institute Work Package 4
Constrained optimization with CAD-free and CAD-based optimization
CAD-based shape optimisation of a S-Bend climate duct of a vehicle
CAD-based and CAD-free aerodynamic optimization of geometrically complex turbomachinery components
Efficient Automatic Differentiation of CAD systems University of Paderborn Work Package 4
Utilizing morphing modules developed in IODA Work Package 2
Constrained and deformable patches securing a smooth transition Work Package 5