Emerging Technology in Optimization An Image Based Approach for CAE

Transcription

1 OptiCon 98 Emerging Technology in Optimization An Image Based Approach for CAE Noboru Kikuchi

2 Major Collaborators Alejandro Diaz Michigan State University Scott Hollister University of Michigan and Keizo Ishii QUINT Corporation

3 Graduate Students in CML current Emilio Silva Shinji Nishiwaki & Susumu Ejima J.H. Yoo Bing-Chung Chen Daichi Fujii Minako Sekiguchi

4 CAE at Present An Introduction to Image Based CAE

5 Current Approach in CAE Parametric (Geometry Based) CAD / CAE Standard CAD Software is based on computational geometry by using parametric spline representation to define shape of a structure/domain All of the existing CAD software are geometry based : Pro-E, UNIGRAPHICS, I-DEAS,CATIA,... In FEA, automatic mesh generation methods are also based on parametric representation of geometry

6 Lots of Sophistication and Big Success (2D,3D?) Realization of importance and profitability of Parametric Geometry Based CAD and CAE

7 Industry Standard in CAD Automotive Industry UNIGRAPHICS in GM I-DEAS in FORD CATIA in CHRYSLER Paradigm Shift in 90s Leading companies have given up In- House CAD/CAE software

8 CAD/CAE Acceptance Not Yet 2D CAD is widely accepted, but 3D CAD is too sophisticated for majority of designers and manufacturers CAE becomes an accepted tool for single disciplinary analysis, but not sufficient to create new value except few areas ( crash, forming, etc )

9 MCAE+FCAE=CAE MCAE(Mechanical CAE) FCAE(Fluid CAE) Two separated CAE, Two separated Preprocessing Software, Two separated CAE analysis specialists..difficulty of Integration for Design and Manufacturing

10 Trend in (M)CAE Major Software Houses MSC/NASTRAN,PATRAN,ABAQUS (US,Europe,Japan) ESI/PAMCRASH,PAMSTAMP,COMPOSIC (Europe,Japan,US) Consolidation Others : Swanson/ANSYS, LS/DYNA, ALGOR,..MDI/ADAMS, Linear Nonlinear Impact (Multi-Body) Design Optimization

11 Two Paths for Survival Total Consolidated MCAE/FCAE Analysis(Linear,Nonlinear,Impact,Multi-body), Design Optimization, Simulation of Manufacturing Processes : Total CAE ESI is a typical example : European s Approach MSC may follow : US for survival Integration with (Imbedding to) CAD CAD software absorb linear CAE for Design MCAE is a part of major CAD software

12 CAD Imbedded MCAE CAD absorbs CAE software Simulation of Design Feasibility Based on only Linear Analysis users are Designers rather than Analysts Less Accuracy but user oriented possibly Design Optimization capability DESIGN ORIENTED Short Turn Around Time

13 Effort in MCAE For Shortening of Turn Around Time by Simplifying FE Modeling Methods CAD Linked Automatic Mesh Generation Adaptive FE Methods (h and p elements) Meshless FE Methods (ANALYSIS) Integration with Design Optimization Design Sensitivity Analysis Size,Shape, and Topology Optimization

14 Importance Shortening of Modeling Time Integration of MCAE and FCAE for Design and Simulation of Manufacturing Process PARADIGM CHANGE! Automatic Mesh Generation? How?

15 Image Based CAE Originated From/Based On OPTISHAPE Topology Optimization

16 Topology Design Method Shape and Topology Design of Structures is transferred to Material Distribution Design (Bendsoe and Kikuchi, 1986)

17 TDM : 3D Shaping Truly Three-dimensional shaping of a structure for optimum Without parametric shape definition by splines

18 Closely Related to Rapid Prototype Layer by Layer Operation Link with CAD for pixel operation Utility of STL (SLC ) file

19 Typical Layerd Manufacturing Processes mirror He-Cd Laser Elevator table vat of liquid polymer CO laser 2 powder bed roller mirror heating chamber rollers nozzle Roll of raw material powder storage Foam foundation Stereolithography Selective Laser Sintering Fused Deposition Modeling

20 What we have done at University of Michigan in a DARPA Project? Project MAXWELL Two way communication between image and CAD data for Topology Optimization

21 OPTISHAPE : Material Design A Homogenization Design Method for Topology of Structures and Materials Poisson s Ratio - 0.5

22

23 Image Finite Element Modeling Finite Element Analysis Design Optimization Parametric Geometry CAD & Rapid Prototype

24 Image Manipulation Mosaic Filtering pixel/voxel mesh Gray Scale Image Adjust Level of Gray Scale

25 Pinching Filtering Filtering Operation makes design change

26 Image Algebra for Modeling c isa = 253 or 252, c TRR = 255, 0 < c anat < 252 Initial Scaffold defined by Accomplished in PV-Wave using Where mask

27 Resulted Finite Element Model Scaffold/Bone Image Scaffold/Bone Mesh Done by Dr. Scott Hollister using Voxelcon2.0

28 Image Based CAE Voxelcon : a Derivative of OPTISHAPE CAD/CT/MRI Image Scan or Equivalent Ways Image Based Automated CAE Mesh Generation Construction of Common Model for Multiple Analyses Load/Support Condition FE Analysis Image Based Design & Optimization OPTISHAPE for topology.layout design Rapid Prototype by Layered Manufacturing Simulation of Material Processing (Casting etc)

29 Database : Image Rather than STL files,slc files are considered SLC files are stored as IMAGES images are then compressed 25K/slice x 500=7.5M

30 Image Regenerated

31 FEM Model

32 Femur CT from Visible Human Data

33 Virtual Femur with Nail - Rendering 3D Surface Rendering of femur with nail Only screws show through femur Data ready for mesh generation

34 VOXELCON byproduct of OPTISHAPE

35 VOXELCON for I-DEAS Quint Corporation CAD Model by I-DEAS

36 VOXELCON for I-DEAS (2) 75M Voxel Elements 9.4 M Voxel Elements

37 OPTISHAPE Quint Corporation Topology (NK, A. Diaz) Compliant Mechanisms (NK, S. Nishiwaki) Shape (H. Azekami) Size (H. Miura)

38 Extension of OPTISHAPE Structural Design Static and Dynamic Stiffness Design Control Eigen-Frequencies Design Impact Loading Elastic-Plastic Design Material Microstructure Design Young s and Shear Moduli, Poisson s Ratios Thermal Expansion Coefficients Flexible Body Design

39 New Extension of OPTISHAPE Piezocomposite and Piezoelectric Actuator Design For Creation of New Value

40 Introduction Force Mechanical Energy Displacement Piezoelectric Material Electric potential Electrical Energy Electric charge Examples: Quartz (natural) Ceramic (PZT5A, PMN, etc ) Polymer (PVDF)

41 Applications Pressure sensors accelerometers actuators, acoustic wave generation ultrasonic transducers, sonar, hydrophones etc...

42 Constitutive Equations of Piezoelectric Medium E Tij = cijkls e E S Di = ε ik E + e S kl kij k k ikl kl Elasticity equation Electrostatic equation T ij - stress S kl - strain E k - electric field D i - electric displacement c E ijkl - stiffness property e ikl - piezoelectric strain property ε S ik - dielectric property

43 Topology Design holes PZT holes Topology Optimization Change the topology of microstructure (material) or structure (transducer) Improvement in the performance of piezocomposite materials; design of new kinds of transducers for different applications

44 Many Approaches : MDM Simple: Density Method property E ijkl = p x E 0 ijkl fraction of material in each point Material Design t Structure Design Domain Ω General: Homogenization Method y 2 1 a microstructure y 1 1 b θ Structure Design x 2 A point with material x 1 A point with no material

45 Performance Characteristics 1 Hydrophones (Hydrostatic Mode) Hydrostatic Coupling Coefficient ( d h ): Figure of Merit (d h g h ): Hydrostatic Electromechanical Coupling Factor (k h ): 2 h kh = T E ε 33 sh d g h d d d d h = h = d 2 h T ε 33 d

46 Performance Characteristics 2 Ultrasonic Transducers (Thickness Mode) Electromechanical Coupling Factor (k t ): 2 e33 kt = D S c33ε33 Impedance (Z): Longitudinal Velocity (v t ): Z v t = ρc D 33 c ρ D 33 =

47 Reference unit cell for comparison: 2-2 piezocomposite PZT5A Polymer (Spurr) dh (pc/n) d h dhgh (fpa -1 ) d h g h Ceramic Volume Fraction Ceramic Volume Fraction (Poled in the 3 direction) kh k h kt k t Ceramic Volume Fraction Ceramic Volume Fraction 47

48 2D Piezocomposite Unit Cell ultrasonic transducer Initially Optimized Microstructure Piezocomposite P Z T 5 A polymer P Z T 5 A P Z T 5 A air air P Z T 5 A 2-2 piezocomposite polymer 3 1 k 2 e c ε 33 t = D S Suggested Transducer

49 Improvement Improvement in relation to the 2-2 piezocomposite unit cell: d h : 2.5 times ρ Z v t ( same) d h g h : 4.2 times k t : 1.13 times stiffness constraint: c E 11 > N/m 2

50 2D Piezocomposite Unit Cell hydrophone Initially Optimized Microstructure Piezocomposite P Z T 5 A polymer P Z T 5 A P Z T 5 A air P Z T 5 A polymer 3 1 Suggested Transducer

51 Improvement Improvement in relation to the 2-2 piezocomposite unit cell: d h : 2.8 times ρ Z v t ( same) d h g h : 7.1 times k t : 1.13 times stiffness constraint: c E 11 >8.108 N/m 2

52 Experimental Verification Rapid Prototyping: Stereolithography Technique 10 mm Optimized Transducer 10 mm Reference Transducer

53 Experimental Result polymer part bar of PZT5A Measured Performances d h (pc/n) d h g h (fpa -1 ) k t Reference Optimized (Simulation) (229.) (10556.) (0.66)

54 2D Piezocomposite Unit Cell hydrophone Initially Optimized Microstructure Piezocomposite air PZT5A PZT5A 3 optimized porous ceramic 1 54

55 Improvement Improvement in relation to the 2-2 piezocomposite unit cell: d h : 3. times d h g h : 9.22 times k h : 3.6 times stiffness constraint: c E 33 > N/m 2

56 Piezocomposite Manufacturing Microfabrication by coextrusion technique Theoretical unit cell Fugitive Ceramic 56

57 Ceramic Feedrod Reduction Zone SEM Image Extrudate Crumm and Halloran (1997)

58 Theoretical Prototype 3 1 Measured Performances d h (pc/n) d h g h (fpa -1 ) Solid PZT Optimized (Simulation) (257.) (19000.) 80 µm

59 3D Piezocomposite Unit Cell hydrophone piezoceramic Poled in the z direction z y x 59

60 3D Piezocomposite Unit Cell hydrophone piezoceramic Polarized in the z direction y z x 60

61 OPTISHAPE Compliant Mechanism Design A New Release

62 Structural Flexibility Flexibility can provide higher performance or additional function If we can specify the flexible mode appropriately. Applied force Deformed direction

63 Kinematic Synthesis Lumped compliant mechanism Based on traditional rigid body kinematics Lumped compliance (Pivot) Stress concentration Her and Midha (1986), Howell and Midha (1994), (1996)

64 Continuum Synthesis Distributed compliant mechanism Based on the topology optimization method Ananthasuresh et al. (1994, 1995), Frecker et al. (1997) Sigmund (1995), (1996), Larsen et al. (1996)

65 Flexibility and Stiffness Applied traction t 1 u 1 u 1 Γt 1 Maximize L 2 ( u 1 )= t 2 u 1 dγ Γt 2 t 2 Dummy traction Γt 2 Mutual Mean Compliance (MMC) Flexibility at Γ t 2 Minimize L 1 ( u 1 )= t 1 u 1 Γ dγ t 1 Mean Compliance (MC) Stiffness at Γ t 1

66 Formulation of Mutual Stiffness Slide along the line Applied traction t 1 t 2 Γ t 2 Dummy traction Γ t 1 Deformed shape Minimize L 2 ( u 1 ) = t 2 u 1 Γ dγ t 2 Stiffness at Γ t 2 with respect to t 1

67 Compliant Mechanism Design Kinematic function Maximize (MMC) Flexibility Structural function Stiffness Applied force Mutual stiffness Minimize ( MC) Reaction force + + Constrained Motion

68 Multicriteria Optimization Flexibility Maximize (MMC) Trade off Stiffness Minimize (MC)

69 Multi-objective Functions (1) Typical methods to deal with multi-objective problems The weighting method The ε-constraint method The goal programming method MMC ----> Infinite! Nash s Optimum

70 Multi-objective Functions (2) (1) Single flexibility case (a) Maximize MMC MC (b) Maximize w Log(MMC) -(1-w)Log( MC) Variation=w δmmc MMC -(1-w) δmc MC

71 Multi-objective functions (3) (2) Displacement single flexibility case Minimize MC MMC Constraint (3) Multi-flexibility case Maximize -1/Cf Log( 1/Cs Log( Exp(-Cf MMC)) j Exp(Cs MC)) i

72 Compliant Gripper (1) Unconstrained single flexibility 60 Extended Design Domain D Ω s =20% Design domain Ω s =30%

73 Compliant Gripper (2) Deformed shape Mises stress Extracted image design

74 Torsional Compliant Mechanism 10 Unconstrained single flexibility 10 Extended Design Domain D : Applied force : Direction of deformation Design domain Extracted image design

75 Constrained Compliant Gripper Constrained single flexibility 60 Extended Design Domain D Design domain Constrained case Optimal configurations (Ω s =20%)

76 Unified Design of Structures and Mechanisms Current design approach Large Friction force Unified design approach Strut-type suspension Small Friction force Unified parts Sub frame Large change of chamber angle Small change of chamber angle

77 Multi-flexibility Compliant Multi-flexibility Mechanism (1) (1) (2) 10 Extended Design Domain D (2) (1) 10 Design domain Optimal configurations (Ω s =30%)

78 Multi-flexibility Compliant Mechanism (2) : Applied force : Direction of deformation (1) Deformed shape (2) Deformed shape

79 Flextensional Actuator Design Piezoceramic + Flexible coupling structure Coupling Structure Mechanical Transform Amplify output displacement Change displacement direction Provide stiffness

80 OPTISHAPE Actuator Design A New Capability to be Implemented

81 Examples of Flextensional Actuators: Brass PZT Coupling structure PZT Brass Moonie Cymbal Low-frequency applications are considered (inertia effect is neglected)

82 Maximize output displacement ( u) ( trade-off ) Maximize blocking force t φ 2 Q 1 { } { } Max (mean transduction) u F 2 1 Min body U t 3 F2 { } { } (mean compliance) -F 2 u 3 φ 2 PZT x 3 Q 1 PZT U 3 x 1

83 Example 1 Multilayer actuator (common design) P T P Z T P Z T P Z T Z T Brass Design region considered (1/4 symmetry)

84 Multilayer Configuration piezoceramic displacement output displacement FEM verification: original deformed 84

85 (1/4 symmetry) Design Domain (Brass) Design Domain (Brass) Q Example 2 B u w=0.5 3 PZT 1 FEM verification: Optimal topology deformed output displ. Piezoceramic original 85

86 Example 3 u B (1/2 symmetry) Design Domain (Brass) Q 3 Piezoceramic w=0.9 1 Piezoceramic Optimal topology ( ) Ω s = 25% Image interpretation

87 Structural Optimization in Magnetic Fields Future OPTISHAPE Capability

88 Shape of H-magnet Cross Sectional View A quarter Model for Analysis

89 Optimal Shape for Maximizing Total Potential Energy Iron Copper Air Design Domain Design Domain for Optimization (324 elements) Optimal Shape with 60% Volume Constraint

90 Analysis of the Optimal Shape Vector Potential Flux Density Increase the value of Flux Densities in Design Domain (25-40%) Stabilize the Flux Densities

91 Optimal Shape for Prescribed Uniform Fields (432 element model) Prescribed Bx = Prescribed Bx = -0.18, By = 0.05 Prescribed Bx = Prescribed Bx = By = 0.05 Ave. of x components E E+00 Ave. of y components E E-01 Stand. Dev. of x components E E-01 Stand. Dev. of y components E E-01

92 Optimal Shape of the Design Domain for Prescribed Uniform Fields (3-layer, 432 element model) Prescribed Bx = Prescribed Bx = -0.20, By = 0.05 Prescribed Bx = Prescribed Bx = By = 0.05 Ave. of x components E E+00 Ave. of y components E-01 Stand. Dev. of x components E E-01 Stand. Dev. of y components E-01

93 Research Issue in OPTISHAPE Material Design Optimization Young s & Shear Moduli Poisson s Ratios Thermal Exapansion Coefficients Electro-magnetic Properties

94 Following To Dr. O. Sigmund Technical University of Denmark Jun Ono Fonseca and Bing-Chung Chen

95 Three-Phrase Material Design Artificial material mixing rule E = ρ me + ( 1 m) E ( 1) ( 2) α = mα + ( 1 m) α ( 1) ( 2) Design layout of two solid phases and void simultaneously Possible overlap between two phases when m 1or m 0

96 Benchmarking with existing 2-phase bound E (1) ( 1) ( 1) = 10, ν = 0. 3, α = 10., V = 50% E (2) ( 2) ( 2) = 10., ν = 0. 3, α = 10. 0, V = 50% Good expansion material surrounded by Bad expansion material results in the Worst expansion composite α H = L N M O Q P

97 Negative Expansion in the vertical direction E (1) ( 1) ( 1) = 10, ν = 0. 3, α = 10., V = 25% E (2) ( 2) ( 2) = 10., ν = 0. 3, α = 10. 0, V = 10% Void α H = L NM O QP Re-entrant structure

98 Near Zero Expansion in the Horizontal Direction Almost disconnected in the y Again, very complicated structure in terms of manufacturing p1= 0.41 p2= 0.05 p0= 0.54 E H = 0. 6 L NM α H = L N M O Q P O QP

99 Construction of three-phase material by two stage design Given distribution of phase 1, design phase 2 distribution, excluding the domain occupied by phase 1 Mark phase 2 as exclusion, design phase 1 The final micro-structure should be noncomplex and easy to manufacture. Design phase 2 Mark phase 1 as excluded Design phase 2 Mark phase 1 as excluded

100 Example: Reinforcement Design distribution of phase 1 Given a material with properties to be improved E H = L NM O QP E 1 = 10, ν = 0. 3

101 Find the Optimal Distribution of Reinforcement. Phase 2 distribution of phase 1??? Add the reinforcement in a particular pattern to achieved design goal E 2 = 5, ν = 0. 3

102 The Optimal Distribution of Reinforcement distribution of phase 2 The reinforcement phase is nonoverlapping with the original phase

103 Superimpose the two non-overlapping phases Super-impose the two phases to achieve the design goal ( increased rigidity and negative Poisson s ratio ) E 1 = 10, ν = 0. 3 E H = E 2 = 5, ν = 0. 3 L NM O QP

104 Negative expansion in the horizontal direction distribution of phase 1 E (1) ( 1) ( 1) = 10, ν = 0. 3, α = 10., V = 30% E (2) ( 2) ( 2) = 10, ν = 0. 3, α = 10. 0, V = 25% Void Stretch in the y due to temperature rise Shrink in the x due to Poisson s effect Unusual CTE material must encompass structure-like mechanism

105 Negative expansion in the vertical direction Initial phase 1 distribution ( inverted honeycomb) α H = L NM O QP E (1) ( 1) ( 1) = 5, ν = 0. 3, α = 10, V = 35% E (2) ( 2) ( 2) = 10, ν = 0. 3, α = 5, V = 25% Void

106 Near zero Thermal Expansion Phase 1 Phase 2 α H = L N M E (1) ( 1) ( 1) = 5, ν = 0. 3, α = 10, V = 25% E (2) ( 2) ( 2) = 10, ν = 0. 3, α = 5, V = 30% O Q P

107 Topology Optimization Algorithm Examination / Research Various Filtering Schemes Proposed using SLP

108 Example 1 Design Domain P Va: 10% Va: 20% Va: 30%

109 Example 2 16 Design domain E=100Gpa P=1 P 10 Va:37.5% CE: Va:25% CE: Va:50% CE:

110 Example 3 8 Design domain 5 4 1

111 Maximization of Attractive Force g N i k = ρ ρ 2 r i = 1 4 k = 1 ik max ρ 2 ρ 1 ρ i ρ 3 ρ 4 ρ = 1 α β i i i r ik :distance N : number of element w : weight g = g / g g ini Objective f = C 1 ( w g) g 2

112 Attractive Forces r ik = 1 ρ i = 1 g i = 4 g i = 3 g i = 2 g i = 1 g i = 0 ρ i = 0. 5 g i = 1 g i = g i = 0. 5 g i = g i = 0 ρ i = 0 g i = 0 g i = 0 g i = 0 g i = 0 g i = 0

113 Example 1A w g = 01. Design Domain P Va: 10% Va: 20% Va: 30%

114 w g = Example 2A Design domain E=100Gpa P=1 P 10 Va:37.5% CE: Va:25% CE: Va:50% CE:

115 Example 3A w g = Design domain 5 4 1

116 Gray Scale Penalty D H ( 11, ) D( 11, ) ρ = 1 αβ ( α = β)

117 Example 1B wg = 01., p = 0. 8 Design Domain P Va: 10% Va: 20% Va: 30%

118 Example 2B wg = 01., p = Design domain E=100Gpa P=1 P 10 Va:37.5% CE: Va:25% CE: Va:50% CE:

119 wg = 01., p = 0. 8 Example 3B 8 Design domain 5 4 1

120 Perimeter Control (Muriel BECKERS 1997) p N 4 = l ρ ρ r ik i k i = 1 k = 1 min ρ 2 ρ 1 ρ i ρ 3 ρ 4 ρ = 1 α β i i i l ik :Length of Common Boundary N : number of element w : weight( g), w : weight( p ), g = g / g, p = p / L g p r ini r r Objective f = C 1 ( w g) + ( w p ) 2 2 g p r

121 Perimeter Length ρ i = 1 p ri = 0 p ri = 1 p ri = 2 p ri = 3 p ri = 4 ρ i = 0. 5 p ri = 0 p ri = 0. 5 p ri = 1 p ri = 15. p ri = 2 ρ i = 0 p ri = 4 p ri = 3 p ri = 2 p ri = 1 p ri = 0

122 Example 1C w = 01., p = 08., w = g p Design Domain P Va: 10% Va: 20% Va: 30%

123 wg = 01., p = 08., wp = Example 2C Design domain E=100Gpa P=1 P 10 Va:37.5% CE: The University Va:25% of Michigan, CE: Va:50% CE: Department of Mechanical Engineering

124 Example 3C 8 Design domain w = 01., p = 08., w = g p w = 01., p = 08., w = g p

125 Post Processing of OPTISHAPE Smooth Surface Extruction

126 Example : Caliper

127 OPTISHAPE Mesh From CT Scan 150,000 3-D Elements 9% Weight Reduction

128 Comparison by Sections

129 Interpolation Functions Meshless Approach f ( x) = n c j Φ j ( x ) j =1 where c j = f ( x j ) Φ j ( x ) is defined with non-polynomial function: Φ j ( x ) = a o ( x )w j ( x ) where w j (x ) = w ( x x j ) and w (x ) = exp( αx 2 )

130 Approximation Functions (2) f ( x) = n c j Φ j ( x ) j =1 To Determine let s assume: Φ j ( x ) which yield k-th degree polynomial, Φ j ( x ) = { a o ( x ) + x j a 1 ( x ) + K + x k j a k ( x )}w j ( x) { } a ( x ) o M = 1 K x j k a k ( x ) w j ( x ) f ( x) = f o + f 1 x + K + f k x k Solve for { a o ( x ) K a k ( x )}

131 Approximation Functions (3) a o ( x) M a k ( x) n n w j ( x) L x k j w j ( x ) j =1 j =1 = M O M n n x k 2 j w j ( x) L x k j w j ( x) j =1 j =1 1 1 M x k Recall: Φ j ( x ) = a o ( x )w j ( x ) { } Φ j ( x) = 1 K x j k n n w j ( x) L x k j w j ( x) j=1 j =1 M O M n n x k 2 j w j ( x) L x k j w j ( x ) j =1 j =1 1 1 M x k w j ( x)

132 Reconstruction of a 3-D Model χ Ω h ( x, y, z ) = k max χ h Ω,k ( x, y ) Φ k ( z) k =1 h χ Ω,k Φ k (z) 2D image : Characteristic function of each image Greyscale values (0-255) : Approximation functions Basis Functions

133 Brake Caliper

134 Analysis Result low stress high stress

135 Possible image-based design software

136 Optimization

137 Prototypes

138 Summary Concept of OPTISHAPE : Topology Optimization is continuously extended not only to structures but also materials, mechanisms, electro-magnetic fields, and others

139 VOXELCON for I-DEAS OPTISHAPE for I-DEAS NASTRAN-OPTISHAPE Toward Image Based CAE