Numerical analysis and optimal design for new automotive door sealing with variable crosssection
Zhu Wenfeng ^{n} , Wang Jie, Lin Peijian
College of Mechanical Engineering, Tongji University, Shanghai 201804, China
article info
Article history:
Received 18 July 2013 Received in revised form 27 May 2014 Accepted 30 June 2014 Available online 17 August 2014
Keywords:
Variable crosssection Geometrical optimization Door seal closing effort Finite element analysis
abstract
Automotive door sealing system isolates passenger compartment from water, dust and wind noise. It has the most direct inﬂuences on doorclosing performance, which is determined by crosssection design in terms of its appropriate Compression Load Deﬂection (CLD) property. Traditional sealing structure has uniform geometrical crosssection. It has the shortcomings of bad ﬁtting in corner parts with large curvatures, causing inaccurate doorclosing effort design. Regarding the door panel's complex 3D proﬁle,
numerical analysis and optimal design for new sealing with variable crosssection are developed in this paper. Firstly, the whole sealing is partitioned into several parts. For four nearly straight segments, conventional 2D numerical analysis can still be used to obtain desired geometrical con ﬁguration. For other four curved corner parts with large curvatures, 3D numerical analysis of door closing is applied. Secondly, 2D geometrical crosssection optimization is proposed. Instead of three variables in previous research, ﬁve variables are selected for featuring crosssection geometry and used for next CAD reconstruction with more precision. After comparison between Back Propagation (BP) neural network and the Kriging surrogate model, BP neural network which performs better and efﬁcient in this automotive design optimization ﬁeld is applied for extracting nonlinear mapping between ﬁve cross section parameters and compression load, which were parallely optimized by Genetic Algorithm (GA) and its efﬁciency and accuracy are compared with another evolutionary algorithm of Particle Swarm Optimization (PSO). Thirdly, 3D numerical modeling of four curved corner parts' closing process is realized, of which twisting and bending effects during seal assembly are taken into account, thus minimizing theoretical error and producing more realistic solution. Consequently, the desired geome trical conﬁgurations for both straight parts and corner parts satisfying designated CLD property can be obtained and the whole sealing can be achieved with variable crosssection, resulting in an ideal door closing effort. Finally, a Matlabbased platform has been developed to assist the design and optimization process. Experiment and case study indicates that it provides an effective method for new door sealing design with variable crosssection.
& 2014 Elsevier B.V. All rights reserved.
1. Introduction
Automotive door seals are installed in narrow gaps between door and body frame along the perimeters of the opening panels, as shown in Fig. 1. They prevent water and dust from entering passenger compartment and accommodate metal manufacturing variations [1]. Doorclosing effort is determined by six factors of seal rubber's compression load, cabin volume, door weight, latch, etc. It was revealed that door sealing consumes 35–50% energy during the door closing process [2], thus making it become the dominant role for door closing effort design [3–5].
^{n} Corresponding author. Email address: Zhuwenfeng@tongji.edu.cn (W. Zhu).
0168874X/& 2014 Elsevier B.V. All rights reserved.
As a component intimately associated with human sensibility, automotive door sealing design includes suitable crosssection's geometrical parameters, as well as appropriate rubber material property, which result in the desired Compression Load Deﬂection (CLD) and the ﬁnal doorclosing efforts. Better sealing perfor mance requires both higher reaction force and a wider contact area. However, better doorclosing performance requires the opposite conditions [6]. A good door sealing design needs to satisfy both the sealing quality and the doorclosing performance. Consequently, door seal design is deﬁned as the compromise between these two reciprocal design targets, which have several features of nonlinearity, such as hyperelastic rubber material, rubber's large deformation and rubber –metal contact behavior. Conventionally, doorclosing effort design and measurement have mostly relied on experimental methods. OrdieresMeré
116 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
Fig. 1. Illustration of door sealing.
tested quasistatic driven door forces [7]. Egashira et al. measured doorseal reaction force and calculated the changes in the inner cabin's atmospheric pressure by air ﬂow [8]. However, the narrow and limited closed space between door seal and door panel makes it difﬁcult to monitor the deformation process and hard to obtain the necessary data of compression load by the traditional mea surement methods. Experimental methods have disadvantages in determining the many factors related to door closing effort design and are not particularly feasible for new development of opti mized seal design regarding to cost and time. Nowadays, computational efforts have been implemented for door seal design. For instance, 2D crosssection of the door seal is modeled and analyzed, making extraction of the reaction force and contact area conveniently [9]. Kim et al. developed a numerical process to predict minimum doorclosing velocity and virtual reaction force versus closing time data [10]. 3D doorclosing analysis, using explicit code, was introduced to produce a more realistic solution. Nonlinear ﬁnite element analysis is applied to investigate seal performance, permitting numerous design itera tions to be evaluated quickly prior to manufacturing and testing the ﬁrst prototype parts. Although numerical analysis makes great contributions to the door seal design, there still exists some problems for this high nonlinear behavior. Door seal is generally in the form of dual extrusion bulbs of metalloid sponge and dense rubber. Its mechan ical properties vary with the amount of deformation, previous load history, temperature, frequency and amplitude of the motion [11–13]. In order to predict the accurate door closing performance, seal material properties must be investigated carefully, which is a challen ging task in computational mechanics due to large deformations and the nearly incompressible nature of rubber [14–17]. Moreover, traditional sealing system has uniform geometrical crosssection. Very few parameters of height and wall thickness are selected for geometrical description. It is simple for design, but would cause inaccurate reconstruction of geometrical con ﬁgura tion. Because the crosssection remains unchangeable, it is easy for manufacturing, but has the shortcoming of bad ﬁtting in large curvature corners and cannot satisfy the varied CLD requirements from different locations of door like the roof part, Apillar part, Bpillar part and sill part. Recently, variable crosssection extrusion technologies have been invented. With the computer control of extrusion die, rubber can be manufactured with changed crosssection. With regard to automotive door's complex 3D proﬁle, new sealing with variable crosssection is desired and can be achieved with the help of new extrusion technology. Consequently, more crosssection geome trical parameters are required. Optimal design of new sealing system could beneﬁt greatly from numerical analysis in term of efﬁciency and research cost. Regarding to material nonlinear, contact nonlinear and geometrical nonlinear of door sealing, this
paper provides a practical numerical approach for new automotive door sealing with variable crosssection. It ﬁrstly presented the door sealing partition process based on closing efforts design. A typical door sealing is geometrically divided into eight segments, considering the curvature change of true proﬁle. Secondly, combination of experiment and numerical analysis is used to identify the hyperelastic material model of door seal rubber. Thirdly, it solved the problem of crosssection's geometrical conﬁguration and optimal variable selection. By using BP neural network, the nonlinear mapping mechanism between ﬁve crosssection parameters and the desired CLD is established. Based on Genetic Algorithm, crosssection parameters are paral lely optimized. Fourthly, numerical modeling of 3D door closing is constructed for corner parts, taking into account the inﬂuence of twisting and bending during door seal assembly. By using 2D crosssection optimization for nearly straight segments of door seal and 3D numerical analysis for the rest of the corner parts, the whole sealing can be designed with variable cross section. Finally, a Matlabbased reexploration platform has been developed, assisting this new sealing system design and demonstration.
2. Door sealing partition based on closing efforts design
Door sealing accounts for a substantial portion of door closing effort. Geometrical and topological parameters of seal crosssec tion, as well as the appropriate rubber material property, would result in a desired Compression Load Deﬂection (CLD) and door closing performance in the end. Traditional sealing structure has uniform geometrical crosssection. However, because of door metal panel manufacturing deviation on which seal rubber was mounted and the 3D true corners parts with large curvature, this type of structure has the shortcoming of bad ﬁtting, causing inaccurate doorclosing effort design. New door sealing needs variable crosssection so that it can bring about change in the CLD curve at different locations, thus satisfying the desired different compression loads and seal qualities. Considering the nonplanar proﬁle of practical sealing in engineering, a typical door sealing is partitioned into eight seg ments, of which four are nearly straight segments and four are corner parts with large curvatures, as shown in Fig. 2. The whole sealing compression load required by good door seal performance is then distributed into each part. Since different compression loads could be raised from different crosssectional designs of sealing structure, each of the eight segments' crosssection para meters needs to be optimally designed based on the distributed compression load, so that sum of them could fulﬁll the desired total door closing effort. For four nearly straightly segments, conventional 2D numerical analysis of compression process can still be used because of the unchanged compression direction of the door sheet metal.
Fig. 2. Partition of door sealing system.
W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
117
Fig. 3. Distribution of doorclosing efforts.
However, the 2D crosssection's geometrical conﬁguration of different segments need to be redesigned in order to obtain the assigned different compression loads. For four corner parts with large curvatures, 3D numerical analysis of door closing is applied, considering the bending and twisting effect caused by large curvatures and the preload effect of door sealing assembly process. The whole partition is shown in Fig. 3.
3. Veriﬁcation of sealing rubber material
Door seals use Ethylene Propylene Diene Monomer (EPDM) sponge rubber. It exhibits nonlinear force versus displacement response characterized by hyperelastic models. Different strain energy functions are used to describe the behavior of rubber material, such as the Mooney–Rivlin, Ogden, etc. Experimental and numerical analysis are applied for selection of the appropriate hyperelastic material model. Firstly, uniaxial tension experiments take place, and the data achieved is ﬁtted to the mechanical coefﬁcients of the various material models which are applied to the ﬁnite element analysis. Then the ﬁnite element results are compared with the experi mental ones to obtain the most suitable material model.
3.1. Test of rubber uniaxial extension
Test process is performed based on the standard GB/T 5282009. The dimensions of test specimen are shown in Fig. 4, which also illustrates the plane strain ﬁnite element analysis of test specimen. It is obvious that the middle part of the test specimen with uniform distributed stress is the effective coverage for testing data collection. Three test specimens are stretched at the speed of 500 7 50 mm/min until fracture. The cracked position must be within the effective coverage. The testing process is shown in Fig. 5 and results are shown in Fig. 6.
Fig. 4. Dumbbelllike test specimen.
Fig. 5. Stretching equipment.
Fig. 6. Results of tension test.
Table 1 Coefﬁcient of hyperelastic constitutive equations.
Mooney (2)
Ogden
Foam
Arruda–Boyce Gent
c
_{1}_{0}
¼ 0.02173
μ ¼
0.08792
c _{0}_{1} ¼ 0.21513
α ¼ 3.75166
μ ¼ 0.08837 α ¼ 1.87486 β ¼ 3.81536
μ ¼ 0.04746
μ ¼ 0.1531
λ ¼ 5.937e þ 10
J _{m} ¼ 497636
3.2. Coefﬁcients of the various material models
Table 1 displays the coefﬁcient ﬁtting results of hyperelastic constitutive equations in MSC.Marc, which can be applied in the numerical analysis for the material model veriﬁcation.
118 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
3.3. Hyperelastic material model veriﬁcation
In order to predict sealing performance and doorclosing force, the chemical, mechanical properties and the constitutive model of rubber material must be considered; moreover, the different orders of the same material constitutive model also inﬂuence the reliability of the simulation results. In this study, we utilize the method in Ref. [1] to select suitable material model for the experimental results. Finally, the Foam model is selected in the simulation, as shown in Fig. 7, and its strain energy function is shown in the following equation:
W ¼
N
∑
n ¼ 1
μ
n
α
n
ðλ ^{α} ^{n} þ λ ^{α} ^{n} þ λ
1
2
α n
3
3Þþ
N
∑
n ¼ 1
μ n
β n
ð1 J ^{β} ^{n} Þ
ð1Þ
Among them, μ _{n} , α _{n} and β _{n} are the material constants. λ _{i} are the principal stretch ratios.
4. 2D crosssection optimization for straight segments
In terms of appropriate CLD property according to the distrib uted sealing compression load, geometrical parameters of four straight segments' cross section need to be designed and opti mized. In order to establish the implicit relationship between the section parameters and the CLD property, BP neural network is adopted to extract the highly nonlinear mapping mechanism with the help of Design Of Experiments (DOE) to provide sufﬁcient training samples. Then, genetic algorithm is applied to parallely search cross section parameters.
4.1. Variables for featuring 2D crosssection
Conventional method uses three simple variables of cross section height, rubber thickness and annulus diameter to describe
Fig. 7. Simulation results with various material models.
Fig. 8. Conventional selection of cross section variables.
Fig. 9. Effect of rib structure on the seal CLD performance.
W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
119
seal crosssection. It ignores rubber's nonuniform and irregular shape. Furthermore, it has not taken into account the in ﬂ uence of special convex structure called rib, as shown in Fig. 8 . In fact, rib segment is very sensitive to compression load according to study in this paper, and a penalty function technique is used to handle the contact constraints in the ﬁ nite element contact simulation. Fig. 9 shows the poor sealing performance caused by rib and the comparison of compression response with and without rib structure. It can clearly be seen that rib segment would cause an excessive doorclosing force, meaning that the traditional simple three variables cannot describe crosssection geometry without considering the rib structure. Considering the nonuniform shape and protruding structure, 2D seal section geometry must be reasonably divided into ﬁve parts, as shown in Fig. 10. After crosssection partition, appropriate variables need to be selected for featuring crosssection geometry. Five variables of angle α, angle β , distance d, associated with other two thickness variables, are proposed in this paper, as shown in Fig. 10. It covers the possible changes in term of the thickness, shape and height of the sponge tube. The ﬁve variables can not only ensure the
Fig. 10. Seal cross section partition and variable selection.
portions I, II and III of the sponge tube to be reconstructed with minimum deviation, but also make the crosssection's geometrical design more convenient and accurate.
Table 2 Variable ranges and level deﬁnition.
Part Variable number Variable type Range 
Level 

I 
3 
Angle α 
0–101 
0 
2.5 
5 
7.5 10 

Angle β 
261 7 91 
17 
21.5 26 
30.5 35 

Distance d 
15 7 4 mm 11 
13 
15 
17 
19 

II 
1 
Thickness t _{1} 0.8– mm 
0.8 
1.1 
1.4 
1.7 
2 

III IV 
1 0 
Thickness t _{2} 0.8–2 mm – – 
0.8 
1.1 
1.4 
1.7 
2 
Fig. 12. Comparison of seal crosssection compression experiment and numerical analysis.
Fig. 13. Fixture of seal compression experiment.
Fig. 11. Industrial quality evaluation of the seal CLD.
120 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
4.2. Objective function for crosssection optimization
The result of seal cross section design is its CLD property which can meet the engineering practical requirement, as shown in Fig. 11. The measured CLD for certain 2D crosssection must be located in the upper and low limit zones as shown in the shadow. To evaluate the performance of a seal product under actual conditions, the seal deformation is checked in three critical conﬁgurations depending on the relative position of the door and frames, denoted by D 2, D, D þ 2 in Fig. 11. D represents the complete closing condition, while D 2 and D þ 2 are slightly less and more deformed conditions, respectively. According to the industrial application, the objective function for crosssection optimization can be obtained and, shown in the following equation:
3 
λ _{i} ðF _{i} f _{i} Þ ^{2} 

f _{ð}_{α} ; β ; d; t _{1} ; t _{2} Þ ¼ 
∑ 
ð2Þ 
i ¼ 1
Among them, F _{i} is the compression load of the three key compression displacements; f _{i} is the center values of the deviation range of the seal design requirement; λ _{i} is the weight, in this paper whose f _{1} ¼ 2.5 N, f _{2} ¼ 3.5 N, and f _{3} ¼ 4.5 N. According to the seal energy consumption calculation formula, the relationship of weight between F _{1} , F _{2} , and F _{3} is F _{1} o F _{2} o F _{3} , therefore, λ _{1} ¼ 0.28, λ _{2} ¼ 0.33, and λ _{3} ¼ 0.39.
4.3. BPbased mapping mechanism extraction
Because of the irregular shape and hyperelastic rubber mate rial of door sealing, there exists strong nonlinear mapping mechanism between its input geometrical variables and output compression load, which are difﬁcult to be expressed by the traditional explicit way. BP neural network can realize complicated implicit mapping as long as it has enough hidden layers and points. It was adopted in this paper for extracting mapping between ﬁve variables of crosssection and the desired CLD response with the closest approximation algorithm.
4.3.1. DOEbased training samples establishment
BP neural network needs sufﬁcient training and learning samples. With the Design Of Experiments methods, the input of crosssection parameters is grouped as ﬁve variables and ﬁve levels, as shown in Table 2; orthogonal table is used to arrange the experiment samples and 25 groups of samples are numerically analyzed to obtain the output values of F _{1} , F _{2} , and F _{3} .
Table 3
Comparison of test and simulation results.
Original seal
D 2(2.5 mm)
D(4.5 mm)
D þ 2(6.5 mm)
FEA (N) Experiment (N) Design index (N) e (%)
5.035
5.926
2.5 7 1 15.03
7.083
8.033
4.5 7 1 11.83
8.426
8.626
6.5 7 1 2.32
Table 4
The effect of the hidden layer neurons on neural network.
4.3.2. Numerical analysis of compression load
Seal compression in a narrow gap between door metal panel during doorclosing is characterized as large deformation of hyper elastic material, showing strong geometrical and material non linearity. Finite element analysis aiming at obtaining the training data collection for neural network must do more work in the aspects of element types, material model and boundary condi tions. The results are shown in Fig. 12.
4.3.3. Experiment of crosssection compression
To verify the numerical analysis of crosssection compression, experiment of real door seal compression is performed, as shown in Fig. 13. Since the CLD property is deﬁned by normal compres sion to crosssection, special ﬁxture with upper part and lower part is designed and made. The upper ﬁxture simulated the door
sheet metal, while lower ﬁxture as body frame metal. By adjusting the position of upper ﬁxture and lower ﬁxture, it can ensure that not only the compression load's direction is vertical to cross section, but also the real spatial assembly position in relation to door seal, door panel and frame panel. The results are listed in Table 3. The least error for compression load at D þ 2 between experimental and numerical analysis is 2.32%, which shows a quite good agreement, indicating that numerical results can be used for BP mapping modeling.
4.3.4. Results of mapping mechanism
A three layers BP neural network is established to model the implicit mapping mechanism using large quantity of training samples learning. The neuron numbers of the input and output layer are 5 and 3 respectively. After taking the training function as “trainlm”, the transfer functions as “tansig” and “purelin”, the highest training number is 7000 and the training goal is 1e 7; then each neural network with different hidden layer neurons is made to learn the training samples, whose results are shown in Table 4 (due to the random training process of the neural network, the data in the table is derived from the comparison of 10 training results). The number of the neurons of the hidden layer is chosen as 18, and other parameters are listed as the above content. Fig. 14 shows the convergence curve of the training mean square error of the neural network, and the comparison of the network output and the sample target is also shown in Fig. 15.
Fig. 14. Convergence curve of BP network training error.
Hidden nodes
12
13
14
15
16
17
18
19
Mean square error Training number Convergence to local optimal solution
0.1034
7000
60%
0.0179
7000
60%
9.93e 8
6132
50%
0.0128
7000
60%
9.35e 8
4139
20%
6.87e 8
1134
10%
9.46e 8
1307
10%
9.71e 8
5960
10%
W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
121
Fig. 15. Sample targets and network outputs.
Table 5
Values of theta0, lob and upb.
theta0 
lob 
upb 
[0.08, 0.1, 0.1, 0.02, 0.02] 
[0.01, 0.01, 0.01, 0.01, 0.01] 
[1, 1, 1, 1, 1] 
Table 6
Comparison of the neural network results and kriging ones.
Load Simulated value (N)
Neural network result (N) Kriging result (N)
F _{1} 
5.0353 
4.2255 
4.2256 

F _{2} 
7.0833 
5.9272 
6.3204 

F _{3} 
8.4259 
7.2972 
8.5258 

Table 7 Average ﬁtness of GA and PSO. 

Optimal method 
GA 
PSO 

Average ﬁtness 
0.053394349 
0.053709607 

Table 8 

Operating parameters of genetic algorithm. 

Population Binary places of Crossover 
Variation 
Generation 

size 
variables 
probability P _{c} 
probability P _{m} 
gap G 
40 
20 
0.62 
0.017 
1 
4.3.5. Comparison between Kriging and BP neural network based
on original section The software package Design and Analysis of Computer Experi ments [18] (DACE) is applied to establish a Kriging model, which is a Matlab toolbox for working with Kriging approximations to computer models. There are two important functions in the toolbox, namely daceﬁt and predictor, as shown in the following equations:
½dmodel; perf ¼ dacefitðS; Y ; regr; corr ; theta0; lob; upbÞ ð3Þ
y ¼ predictorðx; dmodelÞ
ð4Þ
The parameter theta0 in daceﬁt is the key to build a ﬁne surrogate model, and there is a principle that when theta is small, the model is relatively smooth while too simple, on the other hand, if theta is too large; the model will change ﬁercely in local
Fig. 16. Curve of the ﬁtness.
Table 9 Best ﬁtness and optimal result.
Variable 
α (deg) β (deg) d (mm) t _{1} (mm) 
t _{2} (mm) 
Best ﬁtness value 

Optimal result 4.39 
17.47 
16.94 
1.11 
1.36 
0.0054 

Original 
4.00 
26.00 
13.86 
1.26 
1.52 
12.0471 
section
Table 10
Comparison of the optimal results and the simulated ones.
Load Requirement Optimal
(N)
value (N)
Simulated
value (N)
(Optimalsimulated)/
simulated (%)
F _{1}
F _{2}
F _{3}
2.5 7 1 3.5 7 1 4.5 7 1
2.120
2.916
4.144
2.187
3.167
4.042
3.1
7.9
2.5
scope and over ﬁt. After lots of trials, we choose the poly0 regression function and the gauss correlation function; with the assumption of anisotropy we choose the following starting point and bounds for theta0, as shown in Table 5. Because of limited samples, in this paper BP neural network is compared with Kriging method by predicting the original section. Table 9 shows the ﬁve variables of the original section and the comparison of BP neural network results and Kriging ones is shown in Table 6. It can be seen that the Kriging results match the simulated values a little better; however it is time consuming to choose a proper value for theta0, lob and upb, which means less efﬁcient than BP neural network.
4.3.6. Comparison between GA and PSO optimization
on crosssection parameters
Genetic algorithm (GA for short), simulating the natural bio logical evolution process and nature genetic mechanism, is a
122 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
Fig. 17. Comparison of original section geometry and optimized one. (a) The section geometry of the optimization result and (b) optimal results of FEA.
Fig. 18. Curvature analysis of the door corner parts.
classical optimization method. Particle swarm optimization (PSO for short), rooting from simulation of swarm of bird, also solves optimization problems [19]. In this paper, GA is compared with PSO through the same surrogate model established by BP neural network. The optimiza tion efﬁciency is evaluated by the average ﬁtness concluded from 50 calculations, and the comparative results indicate that GA can obtain a better average ﬁtness as shown in Table 7, which means that GA performs a little better than PSO.
4.3.7. GAbased crosssection parameters optimization
Genetic algorithm is applied for optimization of ﬁve selected crosssection parameters. The variables classiﬁcation and range
Fig. 19. Finite element model of Bpillar seal assembly.
are shown in Table 2; the ﬁtness function is shown in Eq. (2). After many trials, its best operating parameters can be obtained and are listed in Table 8. Fig. 16 shows the curve of ﬁtness changing during evolution. The comparison of an optimal result and the original section size is shown in Table 9 and its corresponding compression load values are seen in Table 10. Using the GAbased optimal parameters, a new crosssection's geometrical shape is reconstructed precisely and it was again analyzed numerically to obtain corresponding new CLD curve. New crosssection geometry and new CLD were compared to the corresponding old ones and the results are shown in Fig. 17(a). Fig. 17(b) shows the ﬁnite element analytical process and its result is listed in Table 10. It can be seen that the rib structure matches the proﬁle of upper panel perfectly. The contact and seal area are uniformly
W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
123
Fig. 20. Simulated results of the Bpillar seal.
Fig. 21. Dynamic compression load of Bpillar seal.
distributed along the perimeter of upper panel, which greatly increases the seal performance. Three speciﬁc key measurement points of CLD curve, such as D 2, D and D þ 2, before and after are both listed. The errors of the ﬁttest solution optimized by the genetic algorithm are all within 8%, which reaches the require ment of the engineering application. It indicates that the precision of the nonlinear global mapping relationship between the section parameters and the CLD curve, established by BP neural network, is comparatively high.
5. 3D numerical analysis for corner parts
Door panel's proﬁle is 3D nonplanar curve, causing the related door seal to have complex shape and layout. Previous research regarded the whole seal as a single segment. When calculating the total compression load, it just multiplies the CLD results of compression load per unit length with the total length. This simpliﬁcation ignored the complicated interaction of twisting and bending for corner parts with large curvatures. 3D numerical analysis was modeled for calculating four corners' door sealing compression load.
5.1. Large curvature's inﬂuence
Fig. 18 shows door corner parts, of which the curvature was analyzed. It indicates that corner of B pillar has the biggest curvature change. When corner parts were compressed during door closing, the compression load would be generated from both the normal force as well as the tangent force and twisting force, causing inaccurate calculation of compression force if only using the CLD property obtained from 2D crosssection analysis.
5.2. 3D compression model construction
The Bpillar with the biggest curvature is adopted to examine the compression efforts by 3D simulation of doorclosing process. In order to exhibit the distinct nonlinearity of rubber materials and the complex deform behavior caused by contact, Hypermesh is used for meshing and MSC.Marc is applied for nonlinear analysis. Fig. 19 is the ﬁnite element model of Bpillar assembly on door metal panel. It is established with the tetrahedron mesh controlled by the size of 0.6 mm. The door sheet metal and body frame metal are deﬁned as rigid body and the seal is regarded as the deform able body characterized by the constitutive model of Foam after veriﬁcation of sealing rubber material as mentioned. Simulation steps can be described as ﬁxing the pedestal part of the seal, then rotating the body frame metal for 91 around the axis until the door closing position is reached, as shown in Fig. 19. The large curvature of corner proﬁle and changing compression direc tion caused by rotation of door sheet metal are both taken into account, so that large deformation and compression efforts can be simulated accurately.
5.3. Compression load for corner parts
Fig. 20 illustrates the simulated results of the Bpillar seal with length of 82.24 mm. From the results, it can be learned that the direction of seal deformation is gradually changing from one end plane to the other along the installed proﬁle. Meanwhile, the dynamic compression load versus the door closing angle is
124 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
Fig. 22. Design results of variable crosssection seal system.
Table 11
Parameters of variable crosssection seal system.
α (deg) 
β (deg) 
d (mm) 
t _{1} (mm) 
t _{2} (mm) 

Corner1a 
6 
29 
14.96 
1.44 
1.64 
Corner1b 
6 
30 
14.86 
1.55 
1.48 
Straight1 
0 
25 
16.82 
1.87 
1.7 
Corner2a 
5 
16 
13.9 
1.32 
1.3 
Corner2b 
6 
29 
15.3 
1.49 
15.3 
Straight2 
0 
24 
15.12 
1.85 
1.3 
Corner3a 
6 
29 
14.96 
1.44 
1.64 
Corner3b 
6 
29 
14.96 
1.44 
1.64 
Straight3 
0 
24 
15.78 
1.59 
1.27 
Corner4a 
6 
29 
14.96 
1.44 
1.64 
Corner4b 
6 
29 
14.96 
1.44 
1.64 
Straight4 
6 
29 
14.96 
1.44 
1.64 
extracted, as shown in Fig. 21, from which we can learn that the starting angle of contact between the seal and body frame metal is 8.451.
Fig. 23. Door closing effort with angle of 91.
W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
125
Fig. 24. Door closing efforts under wide open.
6. Case study and Matlabbased platform
By using the many steps of compression load distribution, door sealing portion, 2D crosssection optimization for straight parts and 3D numerical analysis for corner parts, the new automotive door sealing with variable crosssection could ﬁnally be achieved. Fig. 22 and Table 11 show the ﬁnal design results. Combined with other 5 factors, like door weights, cabin air pressure, as mentioned above, an advanced integrated Matlab based platform is developed to perform the ﬁnal design and optimization of total door closing effort design. Figs. 23 and 24 show its application for an engineering case study.
7. Conclusion
[1] Five variables are selected for featuring seal crosssection's irre gular geometry, whose nonlinear implicit mapping with required CLD curve can be extracted by BP neural network and ﬁnally be optimized by the combination with Genetic Algorithm. [2] Door panel's large curvature causes sealing, bending and twisting effect on door closing effort. 3D numerical analysis of door corner part assembly and closing process must be modeled so that corner's compression load is calculated directly instead of using CLD from 2D analysis. [3] By door sealing compression load distribution, portion into straight and curved segments, 2D crosssection optimization for straight parts and 3D numerical analysis for corner parts, new sealing with variable crosssection based on complex true proﬁle is developed.
Acknowledgment
This research is supported by the NSFCChina (National Natural Science Foundation of China) Project (No. 51275359), Project of Shanghai Key Laboratory of Digital Manufacture for Thinwalled Structures project (No. 2012005) and Project of TongjiQingpu
CoResearch Platform (201109, 201305). Director Dr. DAI Yuankan and Research Manager XIA Guoyong of HUAYUCOOPER Standard Sealing System CO., LTD. have made lots of contribution for this research project.
References
[1] E. Dikmen, I. Basdogan, Material characteristics of a vehicle door seal and its effect on vehicle vibrations, Veh. Syst. Dyn. 46 (2008) 975–990. [2] A. Kloess, Z.P. Mourelatos, P.R. Meernik, Probabilistic analysis of an automotive bodydoor system, Int. J. Veh. Des. 34 (2004) 101–125. [3] David A. Wagner, Kenneth N. Morman Jr., Yuksel Gur, Nonlinear analysis of automotive door weatherstrip seals, Finite Elem. Anal. Des. 28 (1997) 33–50. [4] Gao Yunkai, Xu Ruiyao, Yang Lei, Ma Junjie, Development of calculation software for automotive side swing door closing energy, Chin. J. Mech. Eng. 23 (2010) 690–698. [5] Zhao Jiancai, Zhou Chixing, Zhu Xunsheng, Analysis of the inﬂuence of the seal structure on door closing force for SANTANA, in: Proceedings of the SAE 2004 World Congress & Exhibition, Reliability & Robust Design in Automotive Engineering, Detroit, Michigan, USA, 2004, http://dx.doi.org/10.4271/2004
011348.
[6] Hyungil Moona, Ho Kim, Sang Bum Kim, Predicted minimum doorclosing velocity based on a threedimensional doorclosing simulation, Finite Elem. Anal. Des. 47 (2011) 296–306. [7] J. OrdieresMere, A. BelloGarcia, V. MunozMunilla, Finite element analysis of the hyperelastic contact problem in automotive door sealing, J. NonCryst. Solids 354 (2008) 5331–5333. [8] E. Yuji, S. Takayuki, U. Ryuichi, K. Tomohiro, N. Seishi, K. Masahito, Develop ment of side door closing effort simulation, Mazda Kibo 17 (1999) 78–84. [9] G.K. Nikas, G. Burridge, R.S. Sayles, Modelling and optimization of rotary vane seals, Proc. Inst. Mech. Eng. Part J—J. Eng. Tribol. 221 (2007) 699–715. [10] H.Y. Kim, H.J. Kim, H. Kim, J.K. Lee, Development of evaluation system for automotive doorseal, in: Proceedings of the KSAE, vol. 2, 2004, pp. 1073–1079. [11] A. Stenti, D. Moens, W. Desmet, Dynamic modeling of car door weather seals: a ﬁrst outline, in: Proceedings of the ISMA 2004 International Conference on Noise & Vibration Engineering, Leuven, Belgium, 2004, pp. 1249–1260. [12] Z.H. Lu, L.R Wang, I. Hagiwara, Finite element simulation of the static characteristics of a vehicle rubber mount, Proc. Inst. Mech. Eng. 216 (2002)
[13] E. Dikmen, I. Basdogan, Modeling viscoelastic response of vehicle door seal, Exp. Tech. 35 (2011) 29–35. [14] A. Stenti, D. Moens, P. Sas, W. Desmet, Development of a numerical modeling methodology for the NVH behaviour of elastomeric line connections, in:
Proceedings of the International Conference on Noise and Vibration Engineer ing, ISMA2006, Louvain, Belgium, 2006, pp. 157–168. [15] J. Bonet, R.D. Wood, Nonlinear Continuum Mechanics for Finite Element Analysis, ﬁrst ed., Cambridge University Press, Cambridge, 1997.
126 W. Zhu et al. / Finite Elements in Analysis and Design 91 (2014) 115 –126
[16] K. Miller, Testing Elastomers for Hyperelastic Material Models in Finite Element Analysis, 2004, Axel Products 2004, 〈http:/www.axelproducts.com/ downloads/TestingForHyperelastic.pdf〉. [17] N.M. Vriend, A.P. Kren, Determination of the viscoelastic properties of elastomeric materials by the dynamic indentation method, Polym. Test. 23 (2004) 369–375.
[18] Lophaven Søren Nymand, Hans Bruun Nielsen, Jacob Søndergaard. DACEA Matlab Kriging toolbox, Version 2.0, 2002. [19] Ioan Cristian Trelea, The particle swarm optimization algorithm: convergence analysis and parameter selection, Inf. Process. Lett. 85 (6) (2003) 317–325.