Parameter based 3D Optimization of the TU Berlin TurboLab Stator with ANSYS optislang

Transcription

1 presented at the 14th Weimar Optimization and Stochastic Days 2017 Source: Parameter based 3D Optimization of the TU Berlin TurboLab Stator with ANSYS optislang Benedikt Flurl Johannes Einzinger ANSYS Germany ANSYS, Inc. November 10, 2016

2 Overview ANSYS, Inc. November 10, 2016

3 Pro and Cons are in opposite Parameter based Optimization Parameter free Optimization ANSYS, Inc. November 10, 2016 Which algorithm is best? For which application?

4 Workbench Framework ANSYS, Inc. November 10, 2016

5 Optimization inside Workbench The Workbench Effect easier to use Easy parametric set up of complex simulations Fully parametric easy use of best praxis automated flows inside Workbench ANSYS, Inc. November 10, 2016

6 Optimization Strategy General Procedure: Design Optimization Gradient Based Genetic Evolutionary Design of Experiments Data Sampling Detecting Correlations Detecting Important Parameters Parameter Space Reduction Response Surface Design Optimization ANSYS, Inc. November 10, 2016

7 Meta-Model of Optimal Prognosis, MoP Basic Points Test Points Linear, quadratic Regression Moving Least Square CoP k k y y ˆy N 2 2 ˆy Y Yˆ k 1 N 1 Y Yˆ Y Yˆ ANSYS, Inc. November 10, 2016

8 Coefficient of Prognosis Optimization Strategy, wrt to CoP Quality of Response Surface Approximation 100% 0% ANSYS, Inc. November 10, 2016

9 Design Optimization Optimization Algorithms: Evolutionary Algorithm Gradient- Based Algorithms Which one is the best? Pareto Optimization Strategy is required! and derived from SA ANSYS, Inc. November 10, 2016 Adaptive Response Surface Genetic Algorithm

10 Meta-Model of Optimal Prognosis, Best-Practice Number of Evaluated Designs? Required Designs=f(#Important Variables, Non-linearity) Check CoP for different number of Designs Numerical Error? Best-Practice CFD! Model Error? Options: Design Optimization Meta-Model in Subspace ANSYS, Inc. November 10, 2016

11 Parametric Geometry Meridional Design Leading Edge Trailing Edge Shroud, r=const ANSYS, Inc. November 10, 2016

12 Trailing Edge Leading Edge Parametric Geometry Blade Design Theta-Angle Beta-Angle Thickness m-coordinate Blade Design on 5 Layers: Blade (Beta) Angles: Bezier-Curve, 5 Control Points Thickness Distribution: Bezier-Curve, 5 Control Points ANSYS, Inc. November 10, 2016

13 Meshing Scalable Block Structured Mesh, automatically smoothed #Control Volumes Min Angle [ ] Volume Ratio [-] ANSYS, Inc. November 10, 2016

14 CFD Set-Up Model: 1 Segment with periodic boundary conditions Material: Air Ideal Gas R = 287 [J/kg/K] c p = 1004 [J/kg/K] Equation System: Mass Momentum Total Energy (+viscous heating) SST Turbulence Model Inlet: Total Pressure= [Pa] Total Temperature= [K] Flow Angle (wrt axis)=42 Outlet: Mass Flow Rate (360 )=9.0 [kg/s] ANSYS, Inc. November 10, 2016

15 CFD, Convergence Study +0.1% % 0.11 Monotonic convergence of all residuals Abs of all imbalances lower than 0.01% Monitor value stationary ANSYS, Inc. November 10, 2016

16 CFD Result - Mach Number Iso Surface: Mach Number= ANSYS, Inc. November 10, 2016

17 CFD Result Losses and Entropy Incompressible Loss Definition: pt in = massflowave(total Pressure)@Inlet pt out = massflowave(total Pressure)@Outlet ps in = massflowave(pressure)@inlet Loss = (pt in pt out)/(pt in ps in) Thermodynamic Loss Definition: s0 = massflowave(static Entropy)@Inlet Entropy = (Static Entropy-s0 )/R Loss S = massflowave(entropy)@outlet Flow Outlet: Flow Angle = atan2(velocity Circumferential,Velocity w) DirOut5 = areaave(flow Angle)@Outlet DirOut4=sum(((Velocity Flow Angle -90[deg])*pi/180[deg])^2)@Outlet sum: wrt to number of Outlet! ANSYS, Inc. November 10, 2016

18 CFD, Mesh Study Δ i [%] = ς i ς fine ς fine incompressible thermodynamic D i Relative Error <1% for 2 Mio Control Volumens Mesh Size for Optimization #Control Volumes ANSYS, Inc. November 10, 2016

19 Parameter Space Input Parameter DeltaBetaXY [-5 ; +5 ] Layer X, Position Y (1-5) DeltaThicknessXY [-0.001; 0] Layer X, Position Y (1-5) Hub Outlet Hub Radius TE Ellipse Ratio, LE/TE Hub&Shroud Layer X Delta wrt to preoptimized 2D design Position Y 4 5 Total 56 Parameter Output Parameter: Loss incompressible thermodynamic Flow Angle ANSYS, Inc. November 10, 2016

20 Sensitivity Analysis Monitoring ANSYS, Inc. November 10, 2016

21 Sensitivity Analysis - Summary CoP = f(#designs) Loss incompressible Loss thermodynamic Flow Angle Conclusion: CoP small increase Important Parameters: small change 2 dominating Parameter per Output N=240 N=240 N=240 N=120 N=120 N=120 Loss incompressible Loss thermodynamic Flow Angle ANSYS, Inc. November 10, 2016

22 Coefficient of Prognosis Sensitivity Analysis next Step 100% NO: CoP too small (<90%) YES: with important or dominating parameters only! 0% YES: in Reduced parameter Space, wrt to dominating parameter! ANSYS, Inc. November 10, 2016

23 Sensitivity Analysis Space Reduction Flow Angle 3 dominating Parameters Min/Max Bounds modified to fulfill Objective Bounds shifted to Area of Interest Loss, incompressible Flow Angle ~ 0 Minimal Loss ANSYS, Inc. November 10, 2016

24 Sensitivity in Sub-Space - Summary CoP = f(#designs) Loss incompressible Loss thermodynamic Flow Angle Conclusion: CoP small increase/decrease Important Parameters: increase High dimensional MoP! N=320 N=320 N=320 N=120 N=120 N=120 Loss incompressible Loss thermodynamic Flow Angle ANSYS, Inc. November 10, 2016

25 Sensitivity in Sub-Space Response Surface Loss incompressible: Full-Space to Sub-Space Visual: refined area, with additional curvature Sub-Space is high dimensional, medium CoP MoP Full-Space: MoP Sub-Space: ANSYS, Inc. November 10, 2016

26 Coefficient of Prognosis Sensitivity in Sub-Space next Step 100% NO: CoP too small (<90%) Use Anyway FAST YES: with important or parameters only! 0% High dimensional MoP high CoP too expensive! ANSYS, Inc. November 10, 2016

27 Loss incompressible min Design Optimization on MoP Pareto Optimization=expensive! Optimization Conflict? Loss min Flow Angle min Yes, but week Optimization on MoP=FAST! Used, even with medium CoP MoP Prediction has medium Quality, due to medium CoP! Flow Angle min ANSYS, Inc. November 10, 2016

28 Loss incompressible min Design Optimization on MoP Quality of Prediction is medium Design Optimization required Weak Conflict between objectives Single Objective Optimization faster Flow Angle min ANSYS, Inc. November 10, 2016

29 Loss incompressible min Design Optimization on CFD Single Objective Optimization> loss min Constraint: Flow Angle < 1 All important Parameter from Sub-Space Meta-Model included Evolutionary Algorithm, Convergence Flow Angle min ANSYS, Inc. November 10, 2016

30 Design Optimization Summary All Results wrt to Parameter Space! Knowledge Gain Meta-Model: Improved Design Area of Interest Important Variables Design Optimization: Further Improvement More efficient due to Important Variables Area of Interest Initial Design Meta-Model Prediction Meta-Model Evaluated Design Optimization D Initial vs Optimized Loss incompr. [%] Loss therm. [1e-3] % % Flow Angle [ ] #Designs ANSYS, Inc. November 10, 2016

31 Design Optimization Result Initial Design Optimized Design ANSYS, Inc. November 10, 2016

32 Design Optimization Outlook Recirculation could not be removed in defined Parameter Space Review Parameter and Limits to avoid recirculation ANSYS, Inc. November 10, 2016

33 Design Optimization Velocity Profile BA=0.24 BA= ANSYS, Inc. November 10, 2016

34 Design Optimization Blade Loading ANSYS, Inc. November 10, 2016

35 Summary & Outlook Summary GUI supported Optimization Process Full Meta-Model delivered: important Parameters Area of Interest Sub-Space Meta-Model Design Improvement Design Optimization Further Improvement Outlook Double check initial: Parameterization Parameter Bounds Forward Convergence Study of Meta-Model Extend Workflow with additional operating Points ANSYS, Inc. November 10, 2016

36 Series of Design points Optimization and HPC Pack License dp4 dp3 dp2 dp1 One set of Solver keys without HPC A lot of calculations! How can these calculations be done in a quick way? + 1 Parametric Pack Unused Cores Four sets of solver keys or One set of solvers and 1 x HPC Parametric Pack 94% reduced time to innovation + 1 Parallel Pack ANSYS, Inc. November 10, 2016