1 Local buckling on large sandwich panels applied to light aviation:

2 Experimental and computation dialogue.

3 M. Ginota,b, C. Bouvetb, B. Castaniéb,*,M. D’Ottavioc, J. Serrab, N. Mahueta

4 a Elixir Aircraft, Rue du Jura, 17000 La Rochelle, France

5 b Institut Clément Ader (ICA), Université de Toulouse, CNRS UMR 5312, INSA, ISAE-Supaéro,

6 INSA, IMT Mines Albi, UPS, Toulouse, France

7 c LEME – EA4416, Université Paris Nanterre, 50, Rue de Sèvres, 92410 Ville d’Avray, France

8 * Corresponding author: [email protected]

9 Keywords: Sandwich structures; Local buckling; Wrinkling; Structural testing; Experimental

10 and computation dialogue; nonlinear FEM.

11 Abstract

12 Wrinkling is a local buckling phenomenon in sandwich structures subjected to compression

13 and shear loading that is challenging for the aircraft design engineer. Numerous wrinkling

14 models are proposed in the literature but historical formulas developed after the Second

15 World War are still widely used by the industry with important knock down factors. Theory-

16 experiment correlation should be the final step in validating and evaluating the models. This

17 article presents an experimental-computational dialogue on structural tests on large

18 sandwich panels of dimensions 558 x 536 mm2 representative of the design used in light

19 aviation. The panels were subjected to compressive and shear loading by using the VERTEX

20 test bench and wrinkling failures were observed. Comparisons are first made with linear
1
21 wrinkling models. Despite correlations are quite encouraging, the imperfection-sensitivity of

22 the experimentally observed wrinkling failure questions the pertinence of a linear bifurcation

23 approach. A nonlinear Finite Element Model (FEM) of the sandwich panels is then developed.

24 Initial imperfections measured by Stereo Digital Image Correlation (SDIC) are directly

25 introduced in the mesh and an elastoplastic constitutive law for the core is implemented.

26 Dynamic explicit computation is used to access the highly nonlinear post-buckling behaviour

27 and matches the test observations very well. The nonlinear FEM provides an improved,

28 conservative prediction of wrinkling loads over the linear models.

29 1 Introduction

30 Wrinkling is a local instability that can occur when a sandwich structure is subjected to

31 compression or shear loading. The length of the wrinkle pattern is of the order of magnitude

32 of the thickness of the sandwich. Wrinkling is considered a primary cause of failure in sandwich

33 structures with thin skins and low core characteristics, such as the configurations used in light

34 aviation. Wrinkling models have been under development since the Second World War and

35 several literature reviews retrace the evolution of wrinkling modelling (Castanié et al., 2020;

36 Ley et al., 1999; Ginot et al., 2021). Without being exhaustive, the first historical formulas were

37 developed with hypotheses using isotropic skins on an elastic foundation in a 2D framework

38 (Hoff and Mautner, 1945; Plantema, 1966; Allen, 1969). More recent works achieved unified

39 models capable of describing global and local buckling modes in sandwich structures (Benson

40 and Mayers, 1967; Leotoing, 2001; Niu and Talreja, 1999; Douville and Le Grognec, 2013).

41 Other models have been enriched by adding orthotropy for the skin (Fagerberg and Zenkert,

42 2005a) and for the core (Vonach and Rammerstorfer, 2000) and the possibility of a multiaxial

2
43 loading (Sullins et al., 1969; Birman and Bert, 2004). Eventually, few authors have proposed

44 higher order formulations with numerical resolutions. An interesting approach is the

45 Sublaminate Generalised Unified Formulation (S.G.U.F) developed by D’Ottavio (2016). Based

46 on a variable kinematic approach, the formulation allows dedicated models to be introduced

47 for skins and core, thus allowing to choose and identify the model required for properly

48 grasping global and/or local response depending on the problem considered (D’Ottavio et al.,

49 2016; Vescovini et al., 2018). To a lesser extent in the literature, experimental-computational

50 dialogues on sandwich structures have been achieved. Thomsen et al. (1996a; 1996b) worked

51 on the influence of the ply drops on sandwich beams with honeycomb core from an enhanced

52 Winkler model. Out-of-plane skin displacement has been solved numerically. Fagerberg (2004;

53 Fagerberg and Zenker 2005a) carried out important test campaigns on sandwich panels, and

54 his results have been used as a reference for the validation of several wrinkling models

55 (Fagerberg and Zenkert, 2005a; Fagerberg and Zenkert, 2005b). At the structure scale, Tuwair

56 et al. (2016) conducted a test on a sandwich panel under four point bending. Correlations

57 were performed with finite-element analyses and analytical models. Tuwair et al. (2016) noted

58 that using analytical approaches to find an exact solution for wrinkling problems may be

59 limited by the assumptions adopted for these methods. This remark is far from new. After a

60 test campaign, Hoff and Mautner (1945) recommended a knock down factor for their

61 analytical solution. This was taken up by Zenkert’s sandwich construction handbook (1997)

62 and NASA’s technical documents of the late 1960s (Sullins et al., 1969). Most of the

63 experimental results on wrinkling show that the classical formulas provide correct trends but

64 with a non-conservative load prediction (Hoff and Mautner, 1945; Norris, 1964). Therefore,

65 the aeronautic industry often uses the knock down factor recommended by Hoff and Mautner

3
66 to cover this complex phenomenon. The main reason why the analytical expressions are not

67 conservative is that the initial imperfections and displacements before buckling are not

68 considered. The quantification of the effect of these imperfections is difficult especially in

69 the structural cases presented in this study. According to Fagerberg and Zenkert (2005a), in

70 the case of non-negligible imperfections (with an amplitude of the order of magnitude of

71 1/10th of the face sheet thickness), the out-of-plane displacement before buckling will be

72 significant and will increase significantly the non-linear behaviour of the structure.

73 Therefore, Finite Element Models (FEM) often provide a more accurate way of taking

74 geometric and/or mechanical non-linearity into account and investigating the post-wrinkling

75 behaviour.

76 Several approaches can be followed to model a sandwich structure by commercial FE

77 packages. The use of an equivalent single layer model by means of shell elements with a

78 composite section is the less expensive approach but it is inapplicable for wrinkling studies

79 because it discards the out-of-plane deformation of the core. A three-layer model should thus

80 be adopted, where the core is modelled by continuum elements in order to retain its out-of-

81 plane response required for wrinkling analysis. The skins may be modelled by shell elements

82 or by continuum elements, the former approach being computationally more advantageous

83 due to the thinness of the skins, which would require too small continuum elements for

84 avoiding excessive element aspect ratios (Léotoing et al., 2002). Convergence studies

85 performed by Léotoing et al. (2002) also recommended to use a mesh density such that each

86 wavelength of the wrinkling pattern is discretized by at least 8 linear elements. For this

87 matter, a preliminary analytical computation of the expected wavelength is of great help to

88 select a first mesh density.

4
 89 In the literature, models are too often compared to textbook cases – either directly via finite

90 element models, with no geometric defects and perfect boundary conditions (Ginot et al.,

91 2021), or via coupon scale tests that are simple to implement but suffer from underlying

92 boundary condition problems (Ley et al., 1999). A preceeding paper (Ginot et al., 2023) on the

93 subject presents wrinkling tests under multiaxial stresses at the plate scale. Five sandwich

94 panels were tested under compressive and shear loading with the VERTEX test bench (Figure

95 1) and wrinkling type failure was observed. The sandwich panel configurations were chosen

96 to be consistent with the sandwich structures used in light aviation (Castanié et al., 2020; Elixir

97 Aircraft). The VERTEX test bench (Figure 1) develops boundary conditions representative of

98 what an aeronautic sandwich panel can undergo in real conditions (Serra et al., 2017a). This

99 experimental campaign provided a large number of valuable data concerning wrinkling of

100 built-up structures. The present paper thus presents an original and enhanced dialogue

101 between experimental testing and computational modelling; the novelty resides in the

102 consideration of the wrinkling phenomenon at the structural scale within an industry-

103 oriented context related to light aviation, which allows to shed light on various aspects

104 influencing this critical failure mode.

105 The paper is organized as follows. First, the material, methods and results of the VERTEX test

106 campaign discussed in (Ginot et al., 2023) are summarised for presenting a self-contained

107 paper. Subsequently, the experimental-computational dialogue is presented in two main

108 parts: (1) Comparisons are made with linear wrinkling models, results and limits being

109 discussed, and (2) an advanced FEM is developed and the nonlinear analysis is compared to

110 the test results. Eventually, conclusions are drawn.

5
111
112 Figure 1: VERTEX test bench and details of a specimen bolted onto the upper part of the
113 central box.

114 2 Materials, method, and test results

115 2.1 Materials and method

116 The operating principle of the VERTEX machine involves four hydraulic actuators used to load

117 a rectangular box structure. The panel under test closes the upper part of the central box

118 (Figure 1). Actuators 1 and 2 can push or pull symmetrically to bend the box structure of the

119 bench, thus locally loading the tested panel in tension or compression. Similarly, actuators 3

120 and 4 can push to twist the centre of the box structure, thus locally loading the tested panel

121 in shear. The specimens used here were sandwich panels with a monolithic peripheral area

122 and an asymmetric sandwich central area including a tapered region. As a result, the edges

123 were monolithic and were drilled to bolt the specimen to the VERTEX test bench (Figure 2).

6
124
125 Figure 2: Overall panel geometry.

126 The specimen was positioned as shown in Figure 1 and bolted on its 4 sides with 128 screws.

127 The external dimensions were 558 x 536 mm², and the sandwich area with the tapered regions

128 was 390 x 390 mm² and about 21 mm thick (core and skins). The materials used for the

129 specimens were a Polymethacrylimide (PMI) foam of density 51 kg/m3 and prepreg

130 carbon/epoxy woven fabric with additional unidirectional prepreg in the F51_D2 sandwich

131 configuration. The nominal stacking sequences are shown in Table 1. A "one-shot/co-cured"

132 process was used in an autoclave. Specimens were tested under either compressive loading

133 or shear loading (Ginot et al., 2023). Three of the five specimens tested in (Ginot et al., 2023)

134 are studied in the present paper.

135 The observable surface was the upper skin; the lower skin faced the inner test bench box

136 structure. To locate wrinkling in the upper skin, an asymmetric geometry was used to create

137 an offset between the load introduction axis and the mean geometric plane of the sandwich

138 structure. It led to a bending moment and induced an additional compressive load in the upper

139 skin and a tensile load in the lower skin. The upper skin was thus more loaded and was liable

140 to buckle.
7
141 Table 1: Specimen stacking sequence in nominal area. Specimen nomenclature is F51_... for 51
142 kg/m3 PMI foam and …_Dx for specified stacking sequence.

Specimen F51_D1 F51_D2 F51_D3

Loading Compressive Compressive Shear
Fabric +/- 45° Fabric +/- 45° Fabric +/- 45°
2x Unidir 0°
Fabric 0°/90° Fabric 0°/90° Fabric +/- 45°
Stacking
PMI foam PMI foam PMI foam
sequence
Fabric 0°/90° Fabric 0°/90° Fabric +/- 45°
Fabric +/- 45° Fabric +/- 45° Fabric +/- 45°
Fabric +/- 45°
143

144
145 Figure 3: Bending effect on asymmetric sandwich structure.

146 To locate the wrinkling in the centre of the panel, a high-density foam frame was used to

147 reinforce the tapered area. For the specimen under compressive loads, the nominal area was

148 260 x 260 mm2 (dotted frames in Figure 4 (a) and (b)).

8
149
150 Figure 4: Skin definition of the specimens.

151 Skins in the tapered area were reinforced (dark brown area in Figure 4). For the specimen

152 under shear loads, the nominal area (dotted frame in Figure 4 (c)) had a 172 x 172 mm²

153 diamond shape so as to have edges perpendicular to the principal stresses at +/-45°. For

154 specimen F51_D2, a 60 mm wide strip of 2X Unidirectional plies 0° ply was added (blue area

155 in Figure 4 (b)) as a typical stacking used in light aviation. The area where the stacking

156 sequence corresponded to Table 1 was called the “nominal area” (light brown inside the

157 dotted frame area in Figure 4 (a) and (c) for specimens F51_D1 and F51_D3; blue area in Figure

158 4 (b) for specimens F51_D2).

9
159 2.2 Instrumentation

160 Stereo Digital Image Correlation (SDIC) with two 5 Mpx cameras was used, and a speckled

161 pattern was made on the upper face of the specimens (Figure 5 (a)). The acquisition frequency

162 was set to two images per second. Vic3D software (Correlated Solutions Inc., Columbia, SC,

163 USA) was used for post-processing. A high-speed camera (7000 fps) was also used to observe

164 a potential explosive failure (Figure 5 (b)). An infrared camera was added to the setup (Figure

165 5 (c)), allowing for possible wrinkling type failure measurements. In the lower skin, “Rosette”

166 gauges were used (the location of the gauges can be found in Ginot et al, 2023, figure 12).

167
168 Figure 5: Images captured by instrumentation cameras.

169 2.3 Results

170 2.3.1 Strain and curvature fields

171 Figure 6 shows in-plane strains (𝜺𝒙𝒙 field of F51_D1 and 𝜺𝒙𝒚 field of F51_D3). All the frames

172 were taken just before failure. A post-treatment is carried out in order to avoid the grained

173 noise characterizing the SDIC recordings in view of a clearer observation of the wrinkling

174 waves, see (Ginot et al, 2023) for more details. The waves of the wrinkling state do not

175 appear clearly and the whole plate remains under compression or shear respectively. The

176 strains are mainly uniform and follow typical expectations. Therefore, the curvature fields,

10
177 i.e. the inverse diameter of the circle locally tangent to the out-of-plane displacement are also

178 shown Figure 6 (𝑪𝒙𝒙 field of F51_D1 and 𝑪𝒙𝒚 field of F51_D3). The waves appear more clearly

179 and are located near the boundaries of the central area for the compression case; For the

180 shear case, waves are extensively present in the whole area of interest, but larger

181 amplitudes are located near the boundaries also.

182
183 Figure 6: Strains and curvatures evolution fields obtained by SDIC just before failure in
184 specimens F51_D1 and F51_D3.

185 2.3.2 Failure scenario

186 For most specimens, the failure behaviour was similar (Figure 7). The specimens failed by

187 wrinkling in the upper skin, thus validating the design of the specimens. One or more

188 wrinkling waves appeared provoking the core to crush locally. The wave eventually spread

189 over the width with a failure of the core due to tension along the thickness direction. This

11
190 lasted about 2 milliseconds starting from the core crushing. For the shear-loaded specimen,

191 the wrinkling wave followed the direction of the compressive principal stress.

12
192
193 Figure 7: Failure scenario for specimens F51_D1, F51_D2 and F51_D3.

13
194 2.4 Compressive and shear strains at failure

195 Table 2 lists the average principal compressive direction and compressive strain at failure of

196 the upper skin of the sandwich panel in the nominal area for each specimen tested. The

197 average compressive strain at specimen failure is measured thanks to SDIC (numerical

198 extensometer) in the nominal area. For specimens under compressive loading, the principal

199 compressive direction is not more than 4° from the x-axis, which can be considered as pure

200 compression. For the specimen under shear loading, the principal compressive direction is

201 about 52° from the x-axis whereas in a pure shear case it should be at 45°. This is due to a

202 tensile component resulting from coupling between torsion and bending in the VERTEX bench.

203 Therefore, the principal strain 𝜀2 is noted instead of the 45° strain. Note that more

204 information on the measurement method and parameters, the calculation of the average

205 compressive or shear strains, are available in Ginot et al, 2023.

206 Table 2: Average strain at failure.

Specimen F51_D1 F51_D2 F51_D3

Loading Compression Compression Shear
Failure type wrinkling wrinkling wrinkling
Principal
compressive 3.9° 3.6° 51.8°
direction
Average
compressive strain 𝜀𝑥𝑥 = -5400 µ strains 𝜀𝑥𝑥 = -3400 µ strains 𝜀2 = -4600 µ strains
at failure
207

14
208 3 Experimental and computation dialogue: Linear bifurcation analysis

209 in a 2D framework

210 A first try to simulate VERTEX test and the wrinkling phenomena is the use of linear models in

211 a 2D plane strain setting subjected to uniaxial compressive loading. Due to the specific

212 geometry of the panel specimen used in the VERTEX bench (see Figure 3), the sandwich

213 section is not uniformly strained in the axial direction: the upper skin is more loaded and is

214 where the wrinkling instability is of interest, whereas the lower skin carries a lower

215 compressive load due to the overall bending of the specimen. Therefore, it is assumed that

216 the experimentally measured wrinkling pertains to a one-sided mode. Thus, the thickness of

217 the core in the analytical models is chosen to be 50mm, instead of 20mm (the thickness of the

218 VERTEX sandwich specimens), in order to prevent any interactions between the two skins and

219 being comparable to a one-sided mode. Moreover, the same layup is attributed to both skins

220 (symmetric sandwich section) and the skins’ thickness and layup is taken to correspond to the

221 upper skin of the VERTEX sandwich panel (Table 1). Finally, since the load introduction is not

222 as simple, it appears more meaningful to compare critical strains of the upper skin rather than

223 critical loads, which also allows a direct comparison with the experimental results reported in

224 Table 2.

225 3.1 Analytical formulations

226 A number of analytical wrinkling formulae are considered to evaluate the critical bifurcation

227 load in the two-dimensional, plane strain setting, see also (Ginot et al., 2021). In many cases,

228 the skin stiffness appearing in the formula is expressed by the longitudinal Young’s modulus.

229 Since composite laminated skins are studied, this property is computed from the flexural

15
 12
230 rigidity in the direction of the compressive load as 𝐸𝑠 𝑓𝑙𝑒𝑥 = 𝐷∗ 3 (𝐷∗ is the inverse of the
11 𝑡𝑠

231 bending stiffness matrix of the laminate). We refer to the original works for the details of the

232 models taken from the literature: (Hoff and Mautner, 1945; Niu and Talreja, 1999; Leotoing,

233 2001; Douville and Le Grognec, 2013). Note that, for Hoff and Mautner’s pioneering formula:

1⁄
234 𝜎𝑐𝑟𝑖𝑡 = 𝑄(𝐸𝑠 𝑓𝑙𝑒𝑥 𝐸𝑐 𝐺𝑐 ) 3, where 𝐸𝑐 and 𝐺𝑐 are the core modulus and the core transverse

235 shear modulus, respectively, 𝑄 is a constant that has been theoretically calculated at 0.91, but

236 Hoff and Mautner recommend a “practical” constant 𝑄 of 0.5 considering that 𝑄 plays the

237 role of a knock down factor. Both values are compared with test results.

238 As in (Ginot et al., 2021), new results obtained by a S.G.U.F. model are also reported. In order

239 to refrain from simplifying assumptions that may introduce theory-driven inaccuracies, a high-

240 order kinematics with through-thickness cubic axial displacement and quadratic transverse

241 displacement is used for the composite skins. So, parabolic transverse shear strains as well as

242 a linear thickness-stretch are retained. The model for the core is quasi-3D with the

243 displacements approximated by through-thickness polynomials of 12th order. The initial stress

244 matrix is computed upon uniformly straining the whole sandwich section (i.e., the core carries

245 a certain amount of initial compressive load also) and by referring to von Kármán non-

246 linearities. A Navier-type solution is adopted for defining the longitudinal periodic pattern,

247 see, e.g., (D’Ottavio and Polit, 2015; D’Ottavio et al., 2016). Since the wavelength of the

248 wrinkling pattern is an input, the actual wrinkling strain is obtained from the minimum

249 eigenvalue among all processed wavelength responses.

250 For the specimen F51_D3 tested in shear, an equivalent uniaxial model is formulated as

251 suggested by Plantema (1966) and Kassapoglou (2010): assuming that the principal

16
252 compression load occurs at 45° (pure shear), the analytical formulations for wrinkling are used

253 upon “rotating” the relevant quantities in the direction of the applied compression. Here, for

254 skins composed of +/-45° fabrics, this is equivalent to switching to 0°/90° fabrics (see the

255 stacking sequence in Figure 8). It is worth noticing that this approach is expected to be

256 conservative as much as a certain stabilising effect introduced by the tensile load in the

257 direction perpendicular to the compression axis is neglected in the present plane strain

258 setting.

259 3.2 Radar comparison graph

260 The results of the considered models are presented in the radar comparison graph of Figure 8

261 in terms of relative percentage differences between the analytical wrinkling strains and the

262 mean strains at specimen failure (given in Table 2). The dotted line indicates perfect

263 correlation with the tests. A positive percentage shows an optimistic model, a negative

264 percentage shows a conservative model. The wrinkling strain of analytical models are

265 obtained from the corresponding load or stress through the equivalent membrane rigidity of

266 the skin. The minimum critical wrinkling strain between symmetric and antisymmetric

267 wrinkling modes is chosen for analysis.

17
268
269 Figure 8: Radar comparison graph between linear models and VERTEX test results.

270 It is first noticed that the correlation for the specimen F51_D3 is of the same order as for the

271 specimens tested mainly in compression. This confirms that the approach of taking only the

272 compression component seems consistent, even if the specimen is not purely loaded in shear

273 (principal compressive direction at 38°, see § 2.4).

274 The results show that the models of Douville and Le Grognec, Niu and Talreja and Hoff and

275 Mautner (with 𝑄 = 0.91) are optimistic while those of Hoff and Mautner’s with the knock-

276 down factor 𝑄 = 0.5 and of Léotoing et al. provide conservative predictions. The quasi-3D

277 analytical S.G.U.F model performs remarkably well for all configurations, yielding at most

278 errors of about 5% on the conservative side. There are many reasons for the analytical

279 expressions to deviate from the experimental measurements. In general, the models from the

280 literature studied here rely on various ad-hoc assumptions that can produce a certain

281 inaccuracy depending on the considered problem.

18
282 The core model is attributed a major role in the discrepancy between the computed and

283 experimental buckling loads. All analytical models (including the S.G.U.F model) adopt a

284 constant linear elastic behaviour; in addition, the models by Niu and Talreja and Douville and

285 Le Grognec rely on the assumption of an isotropic core. In reality, the PMI foam is slightly

286 anisotropic (Young’s modulus E, tangential modulus G and Poisson’s ratio ν are not related

𝑬
287 by Lamé’s formula 𝑮 = 𝟐(𝟏+𝝂)) and it has different properties in tension and compression (see

288 next § 4.4). In the analytical models, however, the modulus has been taken equal to the simple

289 average between tension and compression. Niu and Talreja (1999) have shown that

290 transverse shear effects can be important with short wavelengths. Léotoing et al.’s model

291 shows conservative correlations with the tests, but the discrepancy with the other models

292 raises questions: it is worth recalling that this model discards the core’s axial stiffness (“anti-

293 plane core”) and postulates a simplistic distribution of the shear stress along the core

294 thickness.

295 Further, questions may be asked about the definition of an equivalent Young’s modulus of

296 the composite laminate skin modelling that neglects the membrane/bending coupling terms

297 (matrix B) for specimens F51_D1 and F51_D2,. However, it is worth recalling that the

298 benchmark (Ginot et al., 2021) has shown that, for this type of stacking (similar thickness and

299 stiffness), this coupling is not of major importance in the critical buckling load and should

300 hence not justify such differences in correlations.

301 Finally, the considered analytical bifurcation buckling formulae do not consider any type of

302 initial imperfection, which is known to be a major reason for excessively optimistic failure

303 loads (Ley et al., 1999; Fagerberg and Zenkert, 2005b). Despite the doubtless attractivity of

19
304 such formulae in terms of required computational effort, this drawback questions their

305 applicability for a reliable sizing. So, analytical wrinkling analysis has been enhanced towards

306 initial imperfections in (Kassapoglou et al., 1995; Fagerberg and Zenkert, 2005b). However, its

307 use in pre-design phases may pose difficulties since specific equipment (as DIC) may be

308 required for identifying imperfection amplitudes. Therefore, the classical knock-down factor

309 approach is widely adopted for artificially reducing the theoretical bifurcation loads, as

310 promoted by the early work of Hoff and Mautner. Their knock-down coefficient 𝑄 = 0.5

311 provides indeed conservative loads in all configurations tested, but it may lead to excessive

312 margins of safety that are detrimental for an optimal lightweight structure.

313 Imperfection sensitivity is known to play a relevant role in presence of equilibrium paths with

314 an unstable post-buckling response. This is precisely the case of sandwich wrinkling because,

315 as experimentally observed, the weak core undergoes compressive or tensile failure as soon

316 as local skin indentation occurs. Since slight initial dents have been observed by SDIC at the

317 upper skin (see (Ginot et al., 2023) for more details), the modeling strategy aims to follow

318 the non-linear response including some material and geometrical non-linearities up to

319 failure. This is the objective of the following sections, in which the experimental-numerical

320 dialogue is extended towards a non-linear FEM accounting for the real initial geometry of the

321 tested panels.

322 4 Development of a non-linear finite elements panel model

323 4.1 3D FE panel model definition

324 The geometry of the sandwich specimen panels, with the integration of the reinforcements

325 (core and skins) around the nominal area (Figure 9) is taken into account in this model. The

20
326 monolithic area bolted to the VERTEX test bench is not represented, just a 5 mm wide strip

327 remains: The area modelled by the FEM is thus 400 x 400 mm². ABAQUS S4R elements are

328 used for the skin and monolithic parts and ABAQUS C3D8R elements are used for the core of

329 the sandwich. The skins and the core are assumed to be perfectly bonded and, therefore, they

330 are linked by coincident nodes. Léotoing et al. (2002) recommend a minimum of 4-5 elements

331 per wrinkling half wavelength. The measured wrinkling half wavelengths in the specimens are

332 about 8 mm to 10 mm depending on the sandwich configuration. The horizontal element size

333 is then set to 2 mm in the nominal area. The total number of degrees of freedom is around

334 720000 for the whole model. The computation is performed using dynamic analysis with

335 ABAQUS Explicit. The choice of this analysis is explained in § 4.4.2.

336
337 Figure 9: FE panel model definition.

338 4.2 Boundary conditions imposed by the displacements measured by SDIC

339 In order to avoid excessively idealized boundary and loading conditions, and in absence of

340 a transfer function explicitly linking the actuator forces of the VERTEX bench to the load

341 acting on the specimen, Stereo Digital Image Correlation (SDIC) data are extracted and used

342 to apply a loading path in the numerical model. This approach has been successfully developed

343 and used for the previous experimental and computation dialogues on VERTEX test campaigns

21
344 (Serra et al., 2017a; Trellu et al., 2020). It consists of imposing the displacements (U,V,W)

345 measured by SDIC on a rectangular frame of upper face nodes (red in Figure 10), plus out-of-

346 plane displacements (W) on 3 additional rectangular frames (orange in Figure 10).

347
348 Figure 10: SDIC loading condition imposed in the FEM by four imposed displacement frames.

349 Rotations are then introduced by the out-of-plane displacement gradient between the loading

350 frames (Sztefek and Olsson, 2008). Displacements are implemented in the model as

351 amplitudes against time and are linearly interpolated between two successive SDIC

352 measurements to represent their evolution during the test.

353 4.3 Initial imperfections

354 4.3.1 Mesh building based on SDIC data

355 The introduction of measured imperfections from a real part in a finite element analysis has

356 been strongly studied in the Thin-Walled Structures community. The most commonly used

357 method to approximate geometrical imperfections is using a double Fourier series (Arbocz,

358 1982). In addition, this formulation allows the Fourier coefficients to be random variables for

359 probabilistic analysis (Wagner et al., 2020). Recently, the use of double Fourier for mesh

360 generation based on SDIC data was performed to model the effects of geometric

22
361 imperfections on the buckling behaviour of woven composite cylindrical shell structures (Xin

362 et al., 2022). In our case, generating imperfections through a double Fourier series is not used.

363 Fourier series such as a sum of periodic sine function would not correctly represent the ply

364 drops offset, present in the sandwich skins between nominal and reinforced areas, because

365 ply drops are not periodic. Another method is to use SDIC measurements of the initial

366 specimen profile and apply them directly to generate a mesh. This method has already been

367 used on CFRP panels under compression (Featherston et al., 2012). A perfect mesh is first

368 made. Then, the out-of-plane positions, z, are extracted from SDIC data at the corresponding

369 x and y positions of the upper skin nodes of the perfect mesh (Figure 11). In nodes at the FEM

370 boundaries, where SDIC data are not consistent or available (SDIC measuring window edges),

371 linear extrapolation from the nearest measured data points is used. The “SDIC” mesh has

372 shown that very small ripples can trigger wrinkling localisation. It is necessary to dissociate the

373 ripples due to the measurement uncertainty from the actual imperfections. The parameters

374 used for the SDIC are recommended by the DIC software to minimise data noises from the

375 image characteristics (resolution, contrast, etc.); Subset size is 35 px; step size is 15 px; with a

376 pixel size equal to 0.2 mm. With these parameters, the average confidence margin of the

377 measurements is about 0.01 pixel, which is equal to 0.002 mm. The imperfections present in

378 the specimens were measured at about 0.04 mm in out-of-plane direction z, which is an order

379 of magnitude greater than the SDIC confidence margin.

23
380
381 Figure 11: Upper sandwich face mesh construction from measured SDIC out of plane
382 coordinates z.

383 Note that the measured surface is the upper face of the skin, the ply drop offsets are

384 integrated into the “SDIC” mesh. The thickness offset in the shell elements of the upper skin

385 is then set to "top" (Figure 12).

386
387 Figure 12: Thickness offset modelled in the shell elements of the upper surface in the “SDIC”
388 mesh.
24
389 4.3.2 Effects of defects

390 Figure 13 illustrates the influence of the initial imperfections introduced into the mesh on in-

391 plane and out-of-plane strain. By comparing the results from a perfect mesh (left image in

392 Figure 13 (a)) and those from the “SDIC” mesh with imperfections (right image in Figure 13

393 (a)), we observe:

394 (1) In Figure 13 (a), wrinkling waves are represented by blue areas with high strain gradients

395 induced by the local bending of the wrinkling waves. In the perfect mesh, the wrinkling pattern

396 is smooth and straight, whereas in the “SDIC” mesh it is less regular and follows the

397 localisation and the geometry of the initial imperfections. The dissymetry of the wrinkling

398 zone displayed in the “perfect mesh” results is induced by the loading which is a pure

399 compressive loading along X axis (Table 2). This dissymetry appears slightly less pronounced

400 in the “SDIC mesh” results. This difference is likely to be due to the non-constant top surface

401 level (z-position due to the geometric imperfections)

402 (2) Figure 13 (b) is the plot of strain evolution of the top and bottom faces in the upper skin of

403 the sandwich at the inspection point (red point in Figure 13 (a)). The Initial imperfections in

404 the “SDIC” mesh induce local bending effects. This is shown by the difference in the in-plane

405 strain evolution between the top and bottom faces (solid lines in Figure 13 (b)). The evolution

406 rapidly becomes non-linear and the buckling onset is characterised by a progressive growth of

407 the non-linearity. In contrast, in the perfect mesh (dotted line in Figure 13 (b)), the strain

408 difference between the top and bottom faces is almost zero and linear until the bifurcation

409 (computation progress equals 1). These local bending effects generate high out-of-plane

410 strains in the core (Figure 13 (c)) from the beginning of the loading. Due to the low mechanical

411 performances of the PMI foam, the out-of-plane strength can be reached rapidly and trigger
25
412 the core to fail. Consequently, a more accurate modelling of the core is proposed in the next

413 subsection.

414
415 Figure 13: FE results of "Perfect" and "SDIC" meshes from the loading condition of specimen
416 F51_D1 (axial compression).”

417 4.4 Core behaviour modeling

418 4.4.1 Constitutive law for the core

419 PMI foam exhibits different mechanical behaviours in compression and tension. In

420 compression, the foam behaves like a ductile material where an elastic response is first

26
421 observed, followed by a plateau corresponding to the buckling of the cell walls and then the

422 cell walls interact, increasing the overall stiffness of the foam (called densification). In tension,

423 the behaviour is rather fragile, with brittle failure. In our case, the difference in mechanical

424 characteristics between compression and tension is significant, with the tension modulus and

425 strength being more than twice the compression ones. Abrate (2008) shows that hydrostatic

426 pressure has an important effect on foam failure. Since the anisotropy in the elastic regime

427 is not very important (Wang et al., 2010), an isotropic bi-modulus constitutive law is

428 assumed and implemented in the principal directions. In compression, a perfect

429 elastoplastic model is used. The foam densification is not modelled here, because it appears

430 for very large strains not reached in our experiments. In tension, an elastic response is used

431 with damage modelled by element deletion (Figure 14). Note that a yield surface with a

432 crushable foam model is available in ABAQUS software, but it does not provide the

433 possibility of working with bi-modulus behaviour. In compression, experimental results

434 indicate that the compressive strength is limited by the buckling of the cell walls and show

435 that it can be closely approximated by a maximum principal stress criterion (Gibson and

436 Ashby, 1997). So, a maximum principal stress criterion is chosen as the yield surface in

437 compression and tension (Figure 14). This greatly simplifies the integration of plastic strains

438 𝜀 𝑝̇ which are classically taken normal to the yield surface. An elliptical yield surface such as

439 described by Deshpande and Fleck (2000) or Huo et al. (2022) would introduce a more

440 complex calculation.

27
441
442 Figure 14: Constitutive law and yield surface of the PMI foam core used in the FE panel
443 model. Parameters 𝐸𝑇; 𝐸𝐶; 𝜎0𝑐 ; 𝜎0𝑡 are calibrated from characterisation test campaigns in
444 traction (ASTM C297 (ASTM C297, 2004)) and compression (ASTM C365 (ASTM C365, 2011)).

445 4.4.2 Explicit computation

446 The constitutive law of the foam was first implemented with implicit computations using

447 ABAQUS UMAT. However, the resulting elasto-plastic instability problem led to difficult

448 convergence problems, a difficulty also noticed by Leotoing (2001). The post-buckling

449 behaviour was not reached in implicit computation. The choice of dynamic explicit

450 computation was therefore made, and the constitutive law was implemented with ABAQUS

451 VUMAT. The explicit solver can be relevant for quasi-static computations when these are

452 subject to convergence problems (significant non-linearity, complex contact management). In

453 the case of buckling, the explicit analysis allows the highly nonlinear post-buckling structural

454 response to be followed (Bisagni, 2000). The use of an explicit solver for the solution of a quasi-

455 static problem has some particularities that have to be dealt with (Pinho, 2005; Serra et al.,

456 2016). Relatively to a standard implicit solver, an explicit solver needs very small increments

457 (depending on the size of the element). Thus, analysis usually requires a large number of
28
458 increments, and a considerable numerical displacement speed compared to the actual quasi-

459 static problem. Kinetic energy is then introduced, and a damping system is needed to reduce

460 the dynamic vibration. During these steps, a numerical error can accumulate, and the work of

461 external forces can be converted into energy other than the internal energy, such as kinetic

462 energy, hourglass control energy and damping energy. It must be checked that additional

463 energy is kept at a negligible level by selecting the right numerical displacement speed while

464 keeping a reasonable calculation time. The numerical displacement speed taken is 100 mm/s

465 with a time increment of 10-7 s. On the other hand, increasing the stable time increment

466 prevents unwanted vibrating effects. The ABAQUS function "*MASS SCALING" (Abaqus

467 Analysis user’s Manual) was used to artificially increase the mass of the model. To ensure that

468 changes in the mass and consequent increases in the inertial forces did not alter the solution

469 significantly, a calculation was performed with a numerical displacement speed that was a

470 quarter of the initial one and showed similar results.

471 In the test process, residual strains are induced in the specimen when it is bolted onto the

472 VERTEX test bench (see detail in (Ginot et al., 2023)). This is transcribed in the FEM’s loading

473 conditions with a preload imposed by large displacements in the four displacement frames

474 (Figure 10). It is necessary to dissipate the kinematic energy introduced by the preload as far

475 as possible. The displacements are too large to be handled by the numerical displacement

476 speed without a significant increase in computation time. The solution shown, in Figure 15, is

477 the simulation of three steps of relaxation after the preload step. In these steps, the velocity

478 of each node of the FEM is set to zero to keep the kinematic energy at zero, then left free so

479 that the panel can take its distorted equilibrium form. During these relaxation steps, the

480 imposed displacements are kept constant (the red dotted line in Figure 15 is the average

29
481 displacement X of the left side of the first displacement frame in Figure 10). The test loading

482 conditions are then introduced. This solution has already been used to dissipate energy after

483 impact in (Serra et al., 2021).

Node veloci es are set to ero

preload relax 1 relax 2 relax Test load

484
485 Figure 15: Overall kinematic energy versus time in FEM from the loading condition of
486 specimen F51_D1. 3 relaxation steps allow kinetic energy dissipation from large
487 displacements in preload step.

488 4.4.3 Effects of the compressive elastoplastic and tensile damage core behaviour

489 The study of the core non-linear behaviour in the local instability is not completely original. In

490 the literature, several authors have tried to add this parameter to their numerical models but

491 have remained at the beam scale. Stiftinger and Rammerstorfer (1997) included a core

492 crushing behaviour in a sandwich beam FEM and reported a slight decrease of the maximum

493 load and a sudden drop in the load. Léotoing et al. (2002) used a perfect elastoplastic model

494 with a von Mises yield function, where the constitutive law is calibrated from a uniaxial

30
495 compression test. They noted that the plastic strains are located where wrinkling occurs. The

496 transition from the elastic to the plastic state was immediately followed by a drastic decrease

497 in the overall stiffness of the sandwich beam In our built-up structure, inclusion of the

498 compressive elastoplastic and tensile damage behavior of the core allows to reproduce the

499 observed failure scenario. Firstly, plastic strains develop locally at the skin interface, then,

500 large out-of-plane strains make the core fail in tension (Figure 16).

31
501
502 Figure 16: FEM cross sections with elastoplastic+damage constitutive law for the core. Plastic
503 strains develop locally, then the core fails in tension (modelled by element deletion).

32
504
505 Figure 17: FE results of elastic and elastoplastic+damage core modelling from loading
506 condition of specimen F51_D1 (compression). The constitutive law introduces an elastoplastic
507 behaviour in compression and an elastic response with damage modelled by element
508 deletion in tension (see Figure 14).

509 Figure 17 illustrates the results of a comparison between a bi-modulus full elastic core

510 behaviour (left image in Figure 17 (a)) and one with the elastoplastic and damage behaviour

511 of the core added (right image in Figure 17 (a)). The transition from an elastic to a plastic state

512 tends to develop non-linear, in-plane strains (solid line in Figure 17 (b)) and out-of-plane

513 strains (solid line in Figure 17 (c)), which increase rapidly with loads. The plastic state develops

514 locally in the panel under the wrinkling waves with maximum amplitude, magnifying the non-
33
515 linearity and localising the wrinkling failure (purple area in Figure 17 (a) right image). The jump

516 observed in (Figure 17 (b) and (c)) at the onset of plasticity is due to the implemented

517 perfectly elastoplastic behaviour. The integration of the elastoplastic and damage behaviour

518 of the core makes the panel fail brusquely (considering when the first tensile failure of the

519 core occurs with the computation progress equal to 1), whereas, in the case of purely elastic

520 core behaviour, buckling is progressive and buckling onset is not easy to interpret.

521 5 Non-linear finite element model of panel: Experimental and

522 computation dialogue

523 5.1 Specimens under compression

524 5.1.1 Overall behaviour

525 Figure 18 and Figure 19 show the displacement fields in the upper face for specimens F51_D1

526 and F51_D2 obtained by SDIC versus those computed by the FEM. Good agreement is found

527 between experimental and numerical displacement fields. In-plane and out-of-plane

528 displacement fields computed by FEM are similar to those measured in the nominal area. This

529 confirms that the loading condition methodology developed in § 4.2 is effective to introduce

530 the loading path imposed by the VERTEX test bench in the sandwich specimen.

34
531
532 Figure 18: Displacement fields: comparison between experimental measurements and FEM
533 results of the upper face of specimen F51_D1 under compressive loading. Frames were taken
534 just before the core failed in tension in the FEM (93% of test progress).

35
535
536 Figure 19: Comparison of displacement fields between experimental measurements and FEM
537 results of the upper face of specimen F51_D2 under compressive loading. Frames were taken
538 just before test failure (99% of test progress).

539 The specimen panels bend globally, as a consequence of the asymmetric geometry (see § 2.1,

540 Figure 3 and (Castanié et al., 2002)), and the upper skin is more loaded than the lower skin

541 (Figure 20). The FEM represents the global membrane and bending stiffness of the sandwich

36
542 panels adequately, as shown by the acceptable correlation in the strain of the upper and the

543 lower skins in the centre of the panel (Figure 20). This confirms the relevance of the

544 mechanical and geometrical characteristics used for the simulation. However, the numerical

545 models are slightly less stiff in bending than either of the specimens, as can be also inferred

546 from the higher values of out-of-plane displacements W in Figure 18 and Figure 19. As a

547 result, the computed compressive strains in the bottom skin are smaller due to the higher

548 tensile load induced by the larger global bending of the panel (Figure 20). As already pointed

549 out by Castanié et al. (2002), the bottom skin of the VERTEX panels is very sensitive to load

550 introduction and local ply drops in the transition region, which can explain the discrepancy.

551 Note that the strains introduced by the assembly of the sandwich panel on the test bench

552 can be seen in Figure 20 and is correctly taken into account in the FEM.

553
554 Figure 20: Strain 𝜺𝒙𝒙 in the upper and the lower skins against test progress for specimens
555 F51_D1 and F51_D2.

556

37
557 5.1.2 Local behaviour

558 In Figure 21 (a) and Figure 22 (a) local gradients in ε_xx strain fields, represented by deep blue

559 areas in SDIC measurements as well as in the simulation results, are present at the edge of the

560 nominal area.

561 This reflects a local bending induced by wrinkles. Wrinkles are shown with the profiles of the

562 curves in Figure 21 (b) and Figure 22 (b) for out-of-plane displacement and in Figure 21 (c) and

563 Figure 22 (c) for strain ε_xx on inspect line L0 (see Figure 21 (a) and Figure 22 (a)). In Figure 21

564 (d) and Figure 22 (d), the strain evolutions at inspect points P1 and P2 are firstly linear with

565 respect to the load and then become non-linear. This is a consequence of the bending effect

566 of the wrinkling wave, which appears at around 80% of the failure load. The strain slopes at

567 P1 and P2 (see Figure 21 (d) and Figure 22 (d)) differ from the averaged one E0. This is due to

568 geometric imperfection generating out-of-plane displacements, which affect local in-plane

569 strains. The difference in slopes and non-linearity at the end of the test, which reflects the

570 onset of buckling, are well represented in the simulations for both specimens. For the

571 specimen F51_D1, the location of the wrinkling predicted by the simulation is consistent with

572 that experimentally observed by SDIC. The profiles of the curves of Figure 21 (b) and (c) show

573 a good match between the test (solid blue lines) and the FEM (dashed red lines), even though

574 larger gradients are present in FEM results. Larger gradients show that the simulation is

575 conservative. Wrinkling is quickly followed by tensile failure of the core, which occurs at 93%

576 of test progress (Figure 21 (b)). At the time studied, i.e. at 93% of the test progress, the local

577 non-linearity resulting from the buckling process is well underway in the simulation whereas,

578 in the test, buckling is just starting to occur. The conservatism of the simulation can be

579 explained by the methodology for introducing imperfections. SDIC measurements of the initial

38
580 profile’s upper face are directly applied to generate the mesh, which implies that the

581 measured curvature is constant over the thickness of the skin. For example, a variation in skin

582 thickness will be simulated as a ripple of the whole skin. It is likely that this method introduces

583 imperfections of greater magnitude than the real ones.

584

39
585
586 Figure 21: Comparison of fields and curves between test and FEM for specimen F51_D1.

587

588

40
589
590 Figure 22: Comparison of fields and curves between test and FEM for the specimen F51_D2.

591

592

41
593 For specimen F51_D2, note that a 60 mm wide strip of two Unidir 0° plies is added in the

594 centre of the panel (Figure 22 (a)) and the wrinkling is positioned in this stiffened area, which

595 drains the loads. Several wrinkles appear in the simulation, whereas only one is present in the

596 test (Figure 22 (a)). This might be explained by the fact that the initial geometrical

597 imperfections are less marked here than for specimen F51_D1 (this has already been

598 suggested in section 3) and are not sufficient to localise the buckling to a particular area in the

599 simulation. Nevertheless, a wrinkle is present where wrinkling is observed in the test. The

600 profiles of the curves in Figure 22 (b) and (c) show a very good match between the test (solid

601 blue lines) and the FEM (dashed red lines).

602 5.2 The specimen under shear (F51_D3)

603 5.2.1 Overall behaviour

604 Figure 23 shows the displacement fields measured by SDIC in the upper face of specimen

605 F51_D3 tested under shear loading and enables them to be compared with those computed

606 by the FEM. The comparison is made just before the test failure. Here again, the correlation

607 between experimental and numerical displacement fields is good.

608

42
609
610 Figure 23: Comparison of displacement fields between experimental measurements and FEM
611 results for the upper face of specimen F51_D3 under shear loading. Frames were taken just
612 before test failure (99% of test progress).

43
613 The correlation of the shear strain in the upper and the lower faces in the centre of the panel

614 is acceptable (Figure 24). The FEM represents the global shear stiffness of the sandwich panel

615 adequately.

616
617 Figure 24: Shear strain 𝜺𝒙𝒚 of the centre of the upper and the lower faces against test
618 progress.

619 5.2.2 Local behaviour

620 Several wrinkles were observed in the nominal area. They are shown in Figure 25 (a) by the

621 local gradients in principal compressive 𝜺𝟐 strain fields and, in Figure 25 (b) and (c), by out-of-

622 plane displacement and 𝜺𝟐 strain curve profiles in the wrinkling area. Note that the principal

623 compressive direction is about 38° from the x-axis, which is why the principal compressive

624 strain 𝜀2 is expressed instead of the 45° strain, which would be pertinent in pure shear loading.

625 The simulation is consistent with the localisation of the wrinkles. The observation of several

626 wrinkling waves allows an accurate determination of the half-wavelength of the buckling

627 pattern; it is around 10 mm. The simulation is doubly validated by the very good correlation

628 between (1) the average strain state and (2) the length of the half-wavelength of the buckling

44
629 pattern. As in the compressive tests, the evolution of the principal compressive strains 𝜺𝟐 at

630 inspect points P1 and P2 shows a linear trend versus loading, followed by a non-linear regime

631 at the end (Figure 25 (d)). Again, this can be attributed to the onset of local buckling that

632 occurred at around 80% of the failure load. The simulation follows the wrinkling phenomenon

633 very well, while being slightly optimistic (Figure 25 (d)).

45
634
635 Figure 25: Fields and curves comparison between test and FEM for the specimen F51_D3.

636 5.3 Radar comparison graph

637 In this section, the radar comparison graph shown in § 3.2 is reused with the results from the

638 non-linear FEM.

46
639
640 Figure 26: Radar comparison graph between models and test results.

641 The advanced nonlinear FE panel model provides an improved and conservative prediction of

642 wrinkling loads over the linear models. The integration of initial imperfections in the mesh and

643 a constitutive law for the core is effective to obtain correlation with the experimental results.

644 Nevertheless, the failure load computed by the non-linear FE panel model is quite similar to

645 the critical load computed by the linear S.G.U.F model. In this study, the strength issue and

646 the stability problem are not so different. But a particular effort has been made on the quality

647 of the specimens, where the maximum amplitude of the initial imperfections has been

648 measured at about 10% of the thickness of the sandwiched panel (Ginot et al., 2023). In an

649 industrial application, the imperfections may be greater and the difference between the two

650 approaches may be significant.

47
651 6 Conclusions

652 An exhaustive experimental and computation dialogue on local buckling, named “wrinkling”,

653 in sandwich panels has been performed in this paper. The test results of three sandwich panels

654 bolted on four sides and tested in compression and shear using the VERTEX test bench have

655 been compared to models. Stacking sequences and material properties (orthotropic

656 asymmetric carbon skins and foam core) are in accordance with an industrial application in

657 light aviation.

658 Firstly, linear 2D plane strain models are challenged with test results. Analytical formulas show

659 optimistic correlations with test results. Beyond a framework (3D stress state; skin orthotropy;

660 bi-modulus behaviour in the core) far from the assumptions on which they were based, these

661 models do not take the initial imperfections into account, which can seriously lower failure

662 loads according to the literature. For specimen F51_D3 tested in shear, the correlation is of

663 the same order as for the specimens tested in compression. This demonstrates that models

664 based on uniaxial loading can be used in shear.

665 Then an advanced non-linear Finite Elements Model of the sandwich panels has been

666 developed. The SDIC measurement data allows the initial shape of the upper face of the

667 sandwich panel to be directly integrated into the FE mesh. Little ripples, considered as

668 imperfections, induce a non-linear local response generating out-of-plane displacements that

669 affect local in-plane and out-of-plane strains. This triggers the failure of the core material. This

670 observation leads to a refinement of the core behaviour. A constitutive law for the core is

671 implemented in the principal directions. The law is isotropic with bi-modulus behaviour. A

672 perfect elastoplasticity model is used for compression and an elastic response with damage

48
673 modelled by element deletion is used for tension. Due to convergence issues, the choice of

674 the dynamic explicit computation is made and run using the explicit solver ABAQUS. The

675 integration of initial imperfections in the mesh allows the buckling in the sandwich panel to

676 be localised. This is observed in tests. Moreover, the constitutive law for the core with failure

677 prediction (crushing and tensile failure) allows the strength approach whereas a linear

678 analytical stability approach is too optimistic (10% on average). The evolution of in-plane

679 strains in the wrinkling area with the local buckling onset is well represented in the

680 simulations.

681 The advanced non-linear FEM of the sandwich panels provides a remarkable prediction of

682 wrinkling compared to the experiment results. However, the model is the result of extensive

683 work on mesh construction, loading condition, and nonlinear material modelling with the use

684 of a dynamic explicit solver. This would not be possible without extensive instrumentation of

685 the tests and the measurement fields offered by the SDIC. In this sense, we have moved away

686 from the means and time available to the engineer for his design. Such a model would be

687 difficult to implement in an industrial context. The linear S.G.U.F model (D’Ottavio and Polit,

688 2015), which is much simpler and applicable to industrial design, also correlates well with the

689 test results. A safety approach could be used with knock down factors in the mechanical

690 characteristics of the core and/or the thickness of the skins. Both are preponderant in the

691 occurrence of wrinkling. This approach, in particular on the thickness, is widely used by the

692 industry for the calculation of buckling of shell structures by global finite element models

693 (GFEM). Similarly, Niu and Talreja, (1999) and Douville and Le Grognec, (2013) analytical

694 models can be used with a knock down factor. In general, for analytical models, the

49
695 assumptions used should be carefully checked so that the theoretical framework does not

696 deviate too much from the real one.

697 Declaration of Competing Interest

698 The authors declare that they have no known competing financial interests or personal

699 relationships that could have appeared to influence the work reported in this paper.

700 Acknowledgements

701 This research is part of a CIFRE PhD thesis in collaboration with the aircraft manufacturer Elixir

702 Aircraft (https://elixir-aircraft.com). This work was partially funded by the “Fondation Jean-

703 Jacques et Felicia Lopez-Loreta pour l’Excellence Académique” as part of the VIRTUOSE

704 (VIRTual testing of aerOnautical StructurEs) project (https://websites.isae-

705 supaero.fr/virtuose/). The authors gratefully acknowledge CALMIP (CALcul en MidiPyrénées,

706 (https://calmip.univ-toulouse.fr) for access to the HPC resources and the fast and effective

707 computations it allows. The authors would also like to thank C3Technologies

708 (http://www.c3technologies.fr) for the quality of the sandwich specimens they manufactured.

709 References

710 Abaqus Analysis user’s manual, Dassault Systemes Simulia, V6.12. chap 11.6.1 Mass scaling.
711 http://193.136.142.5/v6.12/books/usb/default.htm?startat=pt04ch11s06aus74.html#u
712 sb-anl-amassscaling
713 Abrate, S., 2008. Criteria for yielding or failure of cellular materials. J. Sandw. Struct. Mater.
714 10, 5–51. https://doi.org/10.1177/1099636207070997
715 Allen, H.G., 1969. Wrinkling and other forms of local instability, in: Neal, B.. (Ed.), Analysis
716 and Design of Structural Sandwich Panels. https://doi.org/10.1016/b978-0-08-012870-
717 2.50012-2
718 Arbocz, J., 1982. Imperfection Data Bank, a Mean To Obtain Realistic Buckling Loads., in:
719 Ramm, E. (Ed.), Buckling of Shells. pp. 535–567. https://doi.org/10.1007/978-3-642-
50
720 49334-8_19
721 ASTM C297, 2004. Standard Test Method for Flexural Properties of Sandwich Constructions.
722 ASTM C365, 2011. Standard Test Method for Flatwise Compressive Properties of Sandwich
723 Cores.
724 Benson, A.S., Mayers, J., 1967. General Instability and Face Wrinkling of Sandwich Plates
725 Unified Theory and Applications. AIAA J. 5(4), 729-739. https://doi.org/10.2514/3.4054
726 Birman, V., Bert, C.W., 2004. Wrinkling of composite-facing sandwich panels under biaxial
727 loading. J. Sandw. Struct. Mater. 6, 217–237.
728 https://doi.org/10.1177/1099636204033643
729 Bisagni, C., 2000. Numerical analysis and experimental correlation of composite shell
730 buckling and post-buckling. Compos. Part B Eng. 31, 655–667.
731 https://doi.org/10.1016/S1359-8368(00)00031-7
732 Castanié, B., Barrau, J.J., Jaouen, J.P., 2002. Theoretical and experimental analysis of
733 asymmetric sandwich structures. Compos. Struct. 55, 295–306.
734 https://doi.org/10.1016/S0263-8223(01)00156-8
735 Castanié, B., Bouvet, C., Ginot, M., 2020. Review of composite sandwich structure in
736 aeronautic applications. Compos. Part C Open Access 1, 100004.
737 https://doi.org/10.1016/j.jcomc.2020.100004
738 D’Ottavio, M., 2016. A Sublaminate Generali ed Unified Formulation for the analysis of
739 composite structures. Compos. Struct. 142, 187–199.
740 https://doi.org/10.1016/j.compstruct.2016.01.087
741 D’Ottavio, M., Polit, O., 2015. Linearized global and local buckling analysis of sandwich struts
742 with a refined quasi-3D model. Acta Mech. 226, 81–101.
743 https://doi.org/10.1007/s00707-014-1169-2
744 D’Ottavio, M., Polit, O., Ji, W., Waas, A.M., 2016. Benchmark solutions and assessment of
745 variable kinematics models for global and local buckling of sandwich struts. Compos.
746 Struct. https://doi.org/10.1016/j.compstruct.2016.01.019
747 Deshpande, V.S., Fleck, N.A., 2000. Isotropic constitutive models for metallic foams. J. Mech.
748 Phys. Solids 48, 1253–1283. https://doi.org/10.1016/S0022-5096(99)00082-4
749 Douville, M.A., Le Grognec, P., 2013. Exact analytical solutions for the local and global
750 buckling of sandwich beam-columns under various loadings. Int. J. Solids Struct. 50,
751 2597–2609. https://doi.org/10.1016/j.ijsolstr.2013.04.013
752 Elixir Aircraft. https://elixir-aircraft.com
753 Fagerberg, L., 2004. Wrinkling and compression failure transition in sandwich panels. J.
754 Sandw. Struct. Mater. 6, 129–144. https://doi.org/10.1177/1099636204030475
755 Fagerberg, L., Zenkert, D., 2005a. Effects of anisotropy and multiaxial loading on the
756 wrinkling of sandwich panels. J. Sandw. Struct. Mater. 7, 177–194.
757 https://doi.org/10.1177/109963205048525
51
758 Fagerberg, L., Zenkert, D., 2005b. Imperfection-induced wrinkling material failure in
759 sandwich panels. J. Sandw. Struct. Mater. 7, 195–219.
760 https://doi.org/10.1177/1099636205048526
761 Featherston, C.A., Eaton, M.J., Holford, K.M., 2012. Modelling the effects of geometric
762 imperfections on the buckling and initial post-buckling behaviour of flat plates under
763 compression using measured data. Strain 48, 208–215. https://doi.org/10.1111/j.1475-
764 1305.2011.00813.x
765 Gibson, L.J., Ashby, M.F., 1997. Cellular solids. Cambridge University Press.
766 https://doi.org/10.1016/0021-9290(89)90056-0
767 Ginot, M., Bouvet, C., Castanié, B., Serra, J., Mahuet, N., 2023. Local buckling on large
768 sandwich panels applied to light aviation: Experimental setup and failure scenarios.
769 Compos. Struct. 304. https://doi.org/10.1016/j.compstruct.2022.116439
770 Ginot, M., D’Ottavio, M., Polit, O., Bouvet, C., Castanié, B., 2021. Benchmark of wrinkling
771 formulae and methods for pre-sizing of aircraft lightweight sandwich structures.
772 Compos. Struct. 273, 114387. https://doi.org/10.1016/j.compstruct.2021.114387
773 Hoff, N.J., Mautner, S.E., 1945. The Buckling of Sandwich-Type Panels. J. Aeronaut. Sci. 12,
774 285–297. https://doi.org/10.2514/8.11246
775 Huo, X., Jiang, Z., Luo, Q., Li, Q., Sun, G., 2022. Mechanical characterization and numerical
776 modeling on the yield and fracture behaviors of polymethacrylimide (PMI) foam
777 materials. Int. J. Mech. Sci. 218. https://doi.org/10.1016/j.ijmecsci.2021.107033
778 Kassapoglou, C., 2010. Design and Analysis of Composites Structures. Wiley.
779 Kassapoglou, C., Fantle, S.C., Chou, J.C., 1995. Wrinkling of composite sandwich structures
780 under compression. J. Compos. Technol. Res. 17, 308–316.
781 https://doi.org/10.1520/ctr10451j
782 Leotoing, L., 2001. Modélisation Du Flambage Global, Local et Interactif Dans Les Structures
783 Sandwich En Compression. PhD Mines de Saint-Etienne, France.
784 https://www.theses.fr/2001EMSE0020
785 Léotoing, L., Drapier, S., Vautrin, A., 2002. Nonlinear interaction of geometrical and material
786 properties in sandwich beam instabilities. Int. J. Solids Struct. 39, 3717–3739.
787 https://doi.org/10.1016/S0020-7683(02)00181-6
788 Ley, R.P., Lin, W., Mbanefo, U., 1999. Facesheet wrinkling in sandwich structures, NASA/CR-
789 1999-208994.
790 https://ntrs.nasa.gov/api/citations/19990017863/downloads/19990017863.pdf
791 Niu, K., Talreja, R., 1999. Modeling of wrinkling in sandwich panels under compression. J.
792 Eng. Mech. 125, 875–883. https://doi.org/10.1061/(ASCE)0733-9399(1999)125:8(875)
793 Norris, C.B., 1964. Short-column compressive cf strength of sandwich constructions as
794 affected by size of cells of honeycomb core materials. For. Prod. Lab. For. Serv. U.S.
795 Dep. Agric. https://doi.org/10.1007/BF00731082

52
796 Pinho, S., 2005. Modelling composites failure laminated of using physically-based failure
797 models. PhD Imperial College London, UK.
798 https://paginas.fe.up.pt/~stpinho/research/past/phd/
799 Plantema, F. J., 1966. Sandwich Construction: The Bending and Buckling of Sandwich Beams,
800 Plates and Shells, John Wiley and Sons.
801 Serra, J., Bouvet, C., Castanié, B., Petiot, C., 2016. Scaling effect in notched composites: The
802 Discrete Ply Model approach. Compos. Struct. 148, 127–143.
803 https://doi.org/10.1016/j.compstruct.2016.03.062
804 Serra, J., Pierré, J.E., Passieux, J.C., Périé, J.N., Bouvet, C., Castanié, B., 2017a. Validation and
805 modeling of aeronautical composite structures subjected to combined loadings: The
806 VERTEX project. Part 1: Experimental setup, FE-DIC instrumentation and procedures.
807 Compos. Struct. 179, 224–244. https://doi.org/10.1016/j.compstruct.2017.07.080
808 Serra, J., Pierré, J.E., Passieux, J.C., Périé, J.N., Bouvet, C., Castanié, B., Petiot, C., 2017b.
809 Validation and modeling of aeronautical composite structures subjected to combined
810 loadings: The VERTEX project. Part 2: Load envelopes for the assessment of panels with
811 large notches. Compos. Struct. 180, 550–567.
812 https://doi.org/10.1016/j.compstruct.2017.08.055
813 Serra, J., Trellu, A., Bouvet, C., Rivallant, S., Castanié, B., Ratsifandrihana, L., 2021. Combined
814 loadings after medium velocity impact on large CFRP laminated plates: Discrete ply
815 model simulations. Compos. Part C Open Access 6.
816 https://doi.org/10.1016/j.jcomc.2021.100203
817 Stiftinger, M.A., Rammerstorfer, F.G., 1997. Face layer Wrinkling in sandwich shells -
818 Theoretical and experimental investigations. Thin-Walled Struct. 29, 113–127.
819 Sullins, R.T., Smith, G.W., Spier, E.E., 1969. Manual for structural stability analysis of
820 sandwich plates and shells. NASA Contract. Reports CR-1457.
821 https://ntrs.nasa.gov/citations/19700004831.
822 Sztefek, P., Olsson, R., 2008. Tensile stiffness distribution in impacted composite laminates
823 determined by an inverse method. Compos. Part A Appl. Sci. Manuf. 39, 1282–1293.
824 https://doi.org/10.1016/j.compositesa.2007.10.005
825 Thomsen, O.T., Rits, W., Eaton, D.C.G., Brown, S., 1996a. Ply drop-off effects in
826 CFRP/honeycomb sandwich panels - Theory. Compos. Sci. Technol. 56, 407–422.
827 https://doi.org/10.1016/0266-3538(95)00145-X
828 Thomsen, O.T., Rits, W., Eaton, D.C.G., Dupont, O., Queekers, P., 1996b. Ply drop-off effects
829 in CFRP/honeycomb sandwich panels - Experimental results. Compos. Sci. Technol. 56
830 56, 423–437. https://doi.org/https://doi.org/10.1016/0266-3538(96)00007-3
831 Trellu, A., Pichon, G., Bouvet, C., Rivallant, S., Castanié, B., Serra, J., Ratsifandrihana, L., 2020.
832 Combined loadings after medium velocity impact on large CFRP laminate plates: Tests
833 and enhanced computation/testing dialogue. Compos. Sci. Technol. 196, 23.
834 https://doi.org/10.1016/j.compscitech.2020.108194

53
835 Tuwair, H., Volz, J., ElGawady, M.A., Chandrashekhara, K., Birman, V., 2016. Modeling and
836 Analysis of GFRP Bridge Deck Panels Filled with Polyurethane Foam. J. Bridg. Eng. 21,
837 04016012. https://doi.org/10.1061/(asce)be.1943-5592.0000849
838 Vescovini, R., D’Ottavio, M., Do io, L., Polit, O., 2018. Buckling and wrinkling of anisotropic
839 sandwich plates. Int. J. Eng. Sci. 130, 136–156.
840 https://doi.org/10.1016/j.ijengsci.2018.05.010
841 Vonach, W.K., Rammerstorfer, F.G., 2000. Wrinkling of thick orthotropic sandwich plates
842 under general loading conditions. Arch. Appl. Mech. 70, 338–348.
843 https://doi.org/10.1007/s004199900065
844 Wagner, H.N.R., Hühne, C., Elishakoff, I., 2020. Probabilistic and deterministic lower-bound
845 design benchmarks for cylindrical shells under axial compression. Thin-Walled Struct.
846 146, 106451. https://doi.org/10.1016/j.tws.2019.106451
847 Wang, J., Wang, H., Chen, X., Yu, Y., 2010. Experimental and numerical study of the elastic
848 properties of PMI foams. J. Mater. Sci. 45, 2688–2695. https://doi.org/10.1007/s10853-
849 010-4250-9
850 Xin, R., Le, V.T., Goo, N.S., 2022. Buckling identification in composite cylindrical shells with
851 measured imperfections using a Multi-DIC method and finite element analysis. Thin-
852 Walled Struct. 177, 109436. https://doi.org/10.1016/j.tws.2022.109436
853 Zenkert, D., 1997. The handbook of the sandwich construction. Engineering Materials
854 Advisory Services.
855

54