4 5 34 33 32 25 20 14 13 12 11 9 8 7 6 5 4 3 2 10 3

ARTICLE IN PRESS

39 38 35 44 43

Engineering Structures

(2002)

www.elsevier.com/locate/engstruct

45

46

47

A combined nite element based soilstructure interaction model for large-scale systems and applications on parallel platforms
M.C. Genes, S. Kocak
Department of Civil Engineering, Cukurova University, Adana 01330, Turkey Received 17 September 2001; received in revised form 4 March 2002; accepted 6 March 2002

48

49

50 51

52

Abstract A nite element (FE) based soilstructure interaction (SSI) model is presented and parallelized for applications on distributed systems. The SSI model is established by combining two methods: the Consistent Innitesimal FE Cell Method (also referred to as the scaled boundary-nite element method) proposed by Wolf and Song (Finite-element modelling of unbounded media, Wiley, England, 1996) for modelling the soil region extending to innity (far-eld), and the standard FE for the nite region (near-eld) and the structure. By using this combined model, a computer program for harmonic and transient analyses of soilstructure systems is coded. The model is investigated by solving various example problems existing in the literature. The results of the model agree with the results presented in the literature for selected problems. The advantages of the model are demonstrated through these comparisons. In order to decrease the computation time and achieve the solution of large-scale problems, the model is parallelized. As a result of this parallel solution, signicant time is saved for large-scale problems. 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Soilstructure interaction; Transient; Impedance; Parallel; Finite element

53 54 55 56 57 58 59 60 61 62

63 64

65 66 67

1. Introduction In soilstructure interaction (SSI) problems the ability to predict the coupled behaviour of the soil and structure is necessary. Therefore, these problems require combined soil and structure models. While structure models are very well established in the literature, soil models involve complicated analysis due to their unbounded nature. The main difculty in modelling the soil region arises from the propagating wave characteristics in the soil medium. The main goal of many related engineering studies is to develop SSI models, which are reliable and easy to implement. There are numerous valuable works in literature, which are focused on the proposition of mathematical models for soil region of the SSI problems. Most of these mathematical models require a solution dened over a computational mesh. For practical purposes, a limited size computational mesh is required.
Corresponding author. Present address: Scientic Computation Research Center, Rensselaer Polytechnic Institute, CII 7309, Troy, NY 12180-3590, USA. Tel.:+1-518-276-6168; fax:+1-518-276-4886. E-mail address: kocakFs@yahoo.com (S. Kocak).

68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83

1 27 26 28 29 31 30 1 2 3 4 1

Therefore, soil region, which extends to innity, is truncated at certain regions called articial boundaries. The waves propagating through the soil medium should be dissipated at the articial boundaries of the computational mesh. This condition somehow has to be satised in the mathematical models. The main challenge in SSI analyses is the representation of energy dissipated at the articial boundaries. Soil region, which in many cases presents a complex and non-linear behaviour, can be discretized using the Finite Element Method (FEM), the Boundary Element Method (BEM) or hybrid models. All of these have advantages and disadvantages depending on the applied problem. FE models require a large-scale mesh to represent the surrounding soil medium (near-eld), which is bounded by the far-eld that is represented by articial boundaries. In numerical modelling of wave propagation problems the presence of articial boundaries introduces spurious reections, which contaminate the solution. This error can be avoided by introducing special boundary conditions, which might be referred to as radiation damping. These special boundary conditions such as

84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

100

101

102

103

104

105

106

0141-0296/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 1 4 1 - 0 2 9 6 ( 0 2 ) 0 0 0 4 2 - 1

4 5 2 1 3 2 1 3 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162

ARTICLE IN PRESS
2 M.C. Genes, S. Kocak / Engineering Structures (2002)

transmitting, non-reecting and silent boundaries absorb the wave energy. The non-reecting viscous boundaries developed by Lysmer and Kuhlemeyer [2] and White et al. [3] have been widely used for various dynamic soil structure interaction problems. These boundaries are only capable of transmitting plane and cylindrical waves and, therefore, must be located far away from the structure or source of dynamic load. This results in largescale FE meshes. Another way of treating these articial boundaries might be introducing innite elements for modelling the far-eld of SSI problems. These elements were proposed by Bettess [4] to obtain solutions to static and steadystate dynamic problems. Since then, many authors, including Medina and Penzien [5], Yang and Yun [6] and Yerli et al. [7] have successfully extended the innite element formulation to model wave propagation problems in an innite medium. Similar to the articial boundary treatment discussed in the previous paragraph, a ne FE mesh is required to dene the near-eld for reliable predictions when innite elements are used along the truncated regions. BEM is a semi-numerical method and requires obtaining the fundamental solutions pertaining to the region, provided that the radiation condition at innity is satised. It has advantages because only surface discretization of the body is required which decreases the spatial dimension by one. Two- and three-dimensional problems involving the dynamic analysis of surface and underground structures have been solved by the BEM. This is evident in the works of Karabalis and Beskos [8], Israil and Banerjee [9] and Von Estorff et al. [10]. Multi-region Boundary Element Method is a different BEM approach proposed by Tanrikulu et al. [11]. Tanrikulu et al. solved two- and three-dimensional layered type problems and underground structure problems under harmonic and transient loading. BEM decreases the computational mesh signicantly. However, it induces an un-symmetric equation system, and its use is limited when non-linearity is of concern. Hybrid numerical schemes combine the FEM for the structure or the near-eld with another numerical or analytical method for the far-eld. The widely used hybrid model is Coupled Finite Element and Boundary Element Method (FEBEM) both in time and frequency domain. Two- and three-dimensional problems involving the dynamic analysis of surface and underground structures have been solved by the FEBEM. This is evident in the works of Von Estorff and Kausel [12], Von Estorff and Firuziaan [13], Yu et al. [14]. Coupled FEBEM analysis exploits the advantages of the FEM and the BEM and forms a relatively small-size computational mesh, yet requires the conversion of system equations into either FE or BE form. Still, for many engineering problems FE based formulations are preferred and have wider applications.

To obtain a simple yet efcient and reliable soil model for far-eld, an FE based model proposed by Wolf and Song, the Consistent Innitesimal FE Cell Method (CIFEM), is used [1,15,16]. CIFEM is a BE-like method that decreases the spatial dimension by one without requiring a fundamental solution [1]. In their recent papers, Wolf and Song [1719] referred to the CIFEM as The Scaled Boundary-Finite Element Method. In this study, however, the method will be addressed as the CIFEM. The method satises the radiation condition and calculates the dynamic response of the region extending to innity at a truncated region [1]. By using the CIFEM, the chosen truncated soil region representing the far-eld might be very close to the near-eld or the super-structure. This differs from other FE based methods where the far-eld should be kept at a signicant distance from the near-eld. Also, the free soil surface, which has to be discretized in the BEM, can either be eliminated or kept to minimum by using CIFEM. This results with a small sized computational mesh. This and other aspects of the CIFEM will be demonstrated by comparisons in the example problems section. However, the CIFEM results in a system of rst-order non-linear ordinary differential equations (ODE) for the dynamic-stiffness matrix with the independent variable frequency (w). In other FE or BEM based algorithms, the system equations can be written in discrete frequency values. Therefore, through the CIFEM, dynamic-stiffness calculations of the innite region are performed through an integration algorithm in a pre-dened frequency interval. Also, in the CIFEM formulations, the radiation condition, which is satised at high frequency values, is presented by introducing a non-dimensional (ND) frequency. Due to this latter assumption, the integration algorithm for the solution of the resulting ODE is started from high frequency values, which requires some extra calculations until the desired frequency level is reached. Hence, for large-scale problems the computation time increases signicantly because in order to reach a specic frequency result, the calculations have to be performed within a frequency interval until that particular frequency is reached. The calculations made throughout the study unveiled that the vast majority of the time is spent for this integration procedure. This shortcoming of the model is eliminated by introducing a parallelized algorithm for the solution of the rst-order non-linear ODE for the dynamic-stiffness matrix of the innite region. Following the substructure method CIFEM and standard FE method are combined for large-scale SSI problems and a computer program is developed. The CIFEM based computer program, implemented in this study, is validated by solving the example problems presented by Wolf and Song [1], for elastic soil region. Then, CIFEM is extended for viscoelastic material behaviour by introducing hysteretic damping. Following, the model is investigated under harmonic and transient loads. For the trans-

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

4 5 2 1 3 2 1 3 219 220

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 3

formation between frequency and time space, Fast Fourier Transformation (FFT) is used [20].

221

2. Combined SSI model CIFEM, proposed by Wolf and Song, is based on the calculation of the dynamic-stiffness matrix dened on the truncated soil region, which extends to innity. The detailed formulation of this method is given in [1]. The computer program, namely SIMILAR, is also available to calculate only the dynamic-stiffness matrix on the interface of the far and the near-eld for elastic soil media [1]. In this study, a computer code is established by using the CIFEM for far-eld and the FEM for neareld and super-structure. CIFEM used in this study is also extended to viscoelastic materials. Here, we will discuss the basics of the method. In this method, a discretized cell along the near- and far-eld interface is introduced (as in Fig. 1(b)) with its interior boundary coinciding with the near-eld and its exterior boundary with the ctitious interface (see Fig. 1(b)). In the gure, O is the similarity centre [1], and ri and re represent the characteristic length of interior and exterior boundaries of the cell, respectively. Here, the unbounded medium with the characteristic length ri is to be analysed to calculate its dynamic-stiffness matrix. Assembling the dynamic-stiffness matrix of the cell (which is straightforwardly determined from its staticstiffness and mass matrices) and the unknown dynamicstiffness matrix of the unbounded medium characterised

222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246

902 903

by length re results in the unknown dynamic-stiffness matrix of the unbounded medium with length ri. Through the use of compatibility and equilibrium equations between the cell and the bounded and the unbounded regions, the relation between the dynamic-stiffness matrices of the faces of the cell are obtained in frequency space in terms of static-stiffness and mass matrices of the cell. Then by taking the limit as cell width goes to zero and using similarity, the formulation leads to a rstorder non-linear ODE for the dynamic-stiffness matrix in the independent variable w as frequency (see Eq. (A.2) in Appendix A). Here, some features of the CIFEM are listed. The method calculates dynamic-stiffness matrix for a discretized region, which extends to innity. It is a BE-like method. However, it does not require a fundamental solution. It is based on the FE formulation and satises the radiation condition at innity while also decreasing the spatial dimension of the problem by one as in BE. The nodes on the interface coincide with those of the FE mesh of the near-eld. The dynamic-stiffness matrices on the interface are calculated in a frequency interval where the frequencies are automatically determined by the solution algorithm, which will be discussed in Section 3. The procedure is exact in the radial direction and converges to the exact solution in the FE sense on the circumferential directions. It is appealing to perform the limit of the cell width analytically, which leads to the consistent formulation. In the combined SSI model the formulation of the dynamic-stiffness matrix of the near-eld and far-eld is derived in frequency domain. The near-eld is discretized by eight-node and twenty-node quadratic isoparametric elements, and the boundary between the structure and the unbounded medium is discretized by three-node and eight-node line and quadrilateral isoparametric elements for 2D and 3D problems, respectively. Thus, the compatibility of the displacements and equilibrium of forces between the near-eld and far-eld are fully satised. Static-stiffness and mass matrices pertaining to above-mentioned nite elements are very well established and can be found in standard FE books [21]. Here, the general aspects of the model are discussed and wellknown formulations are not repeated. Once the dynamicstiffness matrix of the unbounded medium (S ) is calculated, it is then added to the near-eld dynamic-stiffness matrix, in frequency space, as given in Eq. (1) Fs Fi Sss Ssi Sis Sii S Us Ui (1)

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

904 906 905 907 909 908

Fig. 1. Fundamental concept of FEM with the CIFEM (a) combined SSI model (b) representation of CIFEM.

where S, F and U designate the dynamic-stiffness matrix, force and displacement vectors, respectively. The subscripts i and s in Eq. (1) designate the interface and non-interface nodes of the FE mesh, and superscript designates the far-eld (innite medium) as illustrated

296

297

298

299

300

4 5 2 1 3 2 1 3 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319

ARTICLE IN PRESS
4 M.C. Genes, S. Kocak / Engineering Structures (2002)

in Fig. 1(a). Here, it should be noted that the dynamicstiffness matrices of the nite region can be calculated at discrete frequency values. For S , however, which is the dynamic-stiffness matrix of the innite region, the system of ODE given in Appendix A has to be solved. In the formulations, material damping is introduced by the use of correspondence principle by replacing the real shear modulus (G) by the complex modulus (G(1 2iz)), where, z is the hysteretic damping and i is the imaginary number. In order to test the validity of the implemented program, some example problems are solved to check the correctness of S . The results are compared with those given by Wolf and Song [1] for elastic soil media . Further, the combined SSI model, which is obtained by adding a nite region and extending the CIFEM for viscoelastic soil region, is compared with other models in the literature. The comparisons are presented in Section 4.

320

3. Parallelized SSI model The solution of Eq. (1) can be performed for the required frequencies during the calculation of the dynamic-stiffness matrix of the far-eld (S ) via solving the rst-order non-linear ODE given in Appendix A. The dynamic-stiffness matrix calculations of the far-eld are performed at a frequency interval starting from large ND frequency values using a step-by-step method called BulirschStoer algorithm [22,23]. Several 2D and 3D problems are solved under harmonic and transient loading using the computer program, and results are found to match the existing ones in the literature as presented in Section 4. However, the computation time statistics taken from various examples demonstrated that the majority of the time is spent for the calculation of S . For example, for the problem given in Section 4.5, 80% of the computation time is spent for the calculation of S . This example required 24 h computation time on a PENTIUM 2 400 MHz. processor. Because of this extensive time requirement, parallel solution techniques are sought. Parallel algorithms are widely used in engineering applications where there are large-scale or large-timerequiring problems. Widely used parallel algorithms are based on partitioning a computational mesh and then assigning each partitioned mesh to a separate processor, thus reaching a solution using processors concurrently [24,25]. In this study, the time requiring process is the BulirschStoer integration algorithm applied to the interface region. Our time statistics demonstrated that the time required for the solution of near-eld at each frequency step is relatively less than the far-eld. Therefore, in order to decrease the solution time, the integration algorithm, which is the main time requiring part

321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353

of the model, is parallelized for applications on concurrent systems [26]. BulirschStoer algorithm uses rational function for extrapolation and yields more accurate results and faster solutions for ODE. The advantages of the algorithm were demonstrated in Bulirsch and Stoer [23] by comparing it with alternatives such as: RungeKutta, linear multi-step method of AdamsMoultonBashforth and extrapolation with polynomials. The BulirschStoer algorithm self adjusts the step size and consists of six extrapolations before reaching the next calculation step. After performing six extrapolations, an automatic step size correction is done, and the calculations for the next step start. For the parallel solution of the ODE given in Eq. (A.2), a masterslave structure is used on multiple instruction multiple data (MIMD) concurrent systems. In the parallel algorithm, master processor partitions the matrix into submatrices, and slave processors perform the BulirschStoer calculations in parallel. Processors are assigned to the same amount of work by the master to avoid any delays due to unbalanced work. Therefore, an extra dynamic load balancing process is automatically eliminated. Each processor deals with a portion of the resulting matrix (S ). In Fig. 2, a sample matrix partitioned into four sections is illustrated. To keep interprocessor communication to a minimum, the calculated matrices are updated only once at each step for the following step. Symmetry property of S is considered, and the calculations are performed only for the upper portion of the matrix. For message passing, Parallel Virtual Machine (PVM) software is used and the program is tested on WINDOWS and UNIX machines. The time statistics pertaining to those problems are taken and studied in the following section.

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

4. Results and discussions In this section, some example problems solved by the present SSI model are presented and compared with other methods from the literature, and parallel efciency of the model is discussed. As stated before, in the present

389

390

391

392

393

911 912

913 915 914

Fig. 2.

Partitioned matrix for four processors.

917 916

4 5 2 1 3 2 1 3 394 395 396 397 398 399 400 401

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 5

927 928

model innite soil region is dened by line and surface elements in 2D and 3D cases, respectively. In the remainder, in the FE mesh gures, innite regions, which are calculated by CIFEM, are plotted separately in order to demonstrate the interface between nite and innite regions. During the analysis, the matching nodes of nite and innite regions are automatically combined by the program. 4.1. Circular cavity embedded in full-plane To study the correctness of the developed code, inplane motion of circular cavity embedded in a full-plane is considered (see Fig. 3). This problem is solved by Wolf and Song [1] using SIMILAR for elastic soil region. In order to validate the implementation of innite soil region, this problem is solved by the present code. Due to symmetry, only one-quarter of the structure medium interface is discretized with four three-node line elements. In the comparisons, dynamic-stiffness coefcients are used. The dynamic-stiffness coefcients are calculated by S (w)
T

402

403 404 405 406 407 408 409 410 411 412 413

415 414

S (w)

(2)

416 417 418

where represents the rigid motion of the structure medium interface. The dynamic-stiffness coefcients are represented as spring and damping coefcients as S (a0) G(k(a0) ia0c(a0)) (3)
Fig. 4. Dynamic-stiffness coefcient in frequency domain of circular cavity embedded in full-plane. (a) Spring coefcients, (b) damping coefcients.

420 419

929 931 930

421 422 423 424 425 426 427 428 429 430 431

The radius of the cavity is considered in ND form as r 1. The material properties of the soil medium are given in ND form as: shear modulus G 1, mass density r 1, and Poissons ratio n 1 / 3. First, the soil material is taken as elastic, and the problem is solved by the present code, SIMILAR and a BEM based code [27]. The spring and damping coefcients calculated by three codes for elastic soil region (0% damping ratio) are compared in Fig. 4. From the gures, one can see that the present code and SIMILAR give the same results. This validates the present code for innite soil

932

933

935 934

919 920

region. SIMILAR cannot consider material damping. However, in this study CIFEM is extended to soils with material damping. Therefore, in order to compare the results of the present model for viscoelastic materials only BEM results are given for 2, 5 and 10% damping ratios in Fig. 4. Studying the gures, one can see that the present model gives very close results to that of BEM for various damping ratios. 4.2. Rigid strip foundation resting on a half-space To study how far the cell representing the far-eld (innite soil medium) would be located, a massless rigid strip foundation of width 2B and thickness H on a halfspace is considered (see Fig. 5). Here, the near-eld (nite region) and the far-eld (innite region) material properties are chosen to be the same. The constants dening the soil material behaviour are as follows: the Poissons ratio of the soil is taken as n 0.25, in ND form, the shear modulus and the mass density of the soil are taken as Gs 1 and rs 1, respectively. ND compliance values are studied, which are dened as

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

921 923 922 925 924

450

Fig. 3. Circular cavity embedded in full-plane.

451

4 5 2 1 3 2 1 9 337 938

ARTICLE IN PRESS
6 M.C. Genes, S. Kocak / Engineering Structures (2002)

946 947

948 950 949

Fig. 6.
939 941 940 942 944 943

Vertical compliances of the rigid strip foundation.

952 951

954 955

Fig. 5. FE meshes (a) 72 elements for near-eld (b) 0 element for near-eld (only foundation is discretized).

Cvv
452 453

GsCvv, Chh wB / cs

GsChh, Crr

GsBCrr

a0

(4)

454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485

where cs is the shear wave velocity, a0 is the ND frequency, and subscripts v, h and r designate vertical, horizontal and rocking motions, respectively. In order to study the effects of the truncation for the far-eld, the interface between near- and far-eld is located closer to the foundation by simply discarding rows and columns of the FE mesh of near-eld (see Fig. 5(a)). The nal mesh, which is referred to as only foundation in the following diagrams, is illustrated in Fig. 5(b), which corresponds to the simplest mesh where the nite soil region is not discretized, and only foundation and far-eld are discretized. Here, it is interesting to note that the mesh given in Fig. 5(b) is similar to FEBEM meshes where only the surface of the soil region is discretized in BEM analyses. However, using the present model the free soil surface is not discretized. This should be discretized in BEM analyses within a reasonable distance until the truncation is done. This is a signicant advantage of CIFEM over BEM. Especially in 3D problems, by eliminating or decreasing the size of the free soil surface, a signicant amount of computational effort is saved. In CIFEM the free soil surface that has to be discretized can be eliminated or kept to a minimum, which will be demonstrated in the following diagrams. A total of six FE meshes are studied having 72, 50, 32, 18, 8 and 0 eight-node quadrilateral elements for nite region and 24, 20, 16, 12, 8 and 4 three-node line elements for the interface representing the far-eld. The ND vertical, horizontal and rocking compliance values of the foundation obtained by the present model for various meshes are compared with analytical solutions given by Luco and Westman [28] and FE solutions given by

956 958 957

Fig. 7.

Horizontal compliances of the rigid strip foundation.

960 959

Temel [29] in Figs. 68, respectively. In Ref. [29] innite elements are dened to represent the innite soil region, and the mesh used consists of 58 nite elements and eight innite elements for a circular mesh. The number of elements referred to in the diagrams represents

486

487

488

489

490

962 963

964 966 965

Fig. 8. Rocking compliances of the rigid strip foundation.

968 967

4 5 2 1 3 2 1 3 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 7

the number of nite elements used to model the nite region for the present algorithm. As can be seen from Fig. 6 the vertical compliance distribution obtained by the present model illustrates a good match with the others. It is interesting to note that even the simplest mesh, which represents locating the innite soil region just underneath the foundation, gives the compliance values at a reasonable approximation. Studying the compliance values predicted by the present model for different meshes, one can note that the nite region can be kept at a minimum size. As it is seen in Fig. 6, the results of meshes that have 72, 50, 32, 18 elements for near-eld converged to eight elements. For the sake of simplicity, they are not illustrated in the Figs. 7 and 8. The same observation is made for the horizontal and rocking motions, except only foundation solutions of the present model for horizontal and rocking did not predict the innite soil response as well as in vertical case (see Figs. 7 and 8). Studying the gures, one can see that for surface foundation problems, the present model gives reasonable results with a small discretized nite region. Here, it should be noted that the CIFEM is very advantageous for embedded structures considering the similarity phenomenon [1]. Since, for embedded structures the interface is located at a chosen distance from the similarity centre. The advantages of the CIFEM for embedded structures can be seen in Fig. 6 by investigating the good match for large meshes, which represents the interface for embedded structures. 4.3. Rigid strip foundation resting on a layered halfspace A massless rigid strip foundation of width 2B on a layered half-space is studied (see Fig. 9). The half-space consists of two layers where the top layer has thickness D( 4B). The analyses are performed for two different soil material ratios GL / Gs 4 and GL / Gs 0.25 where

subscripts L and s represent layer and half-space, respectively. In the analyses, the Poissons ratio for the whole soil region is taken as n 0.4, and the following ND quantities are used: Gs 1; rL 1; rs 1. For this problem, a nite region between the soil surface and the bottom of the top layer is dened where the FE mesh is truncated at a 4B distance from the foundation. As a special application in this particular problem, CIFEM is used both for the half-space underneath the top soil layer (extending to innity in vertical direction) and for the truncated regions of the top layer (extending to innity in horizontal) provided the related soil materials are introduced. The horizontal foundation impedance values, calculated by the present model, are compared with the BEM solutions given by Tanrikulu [30] in Figs. 9 and 10 for different soil properties. The impedance and frequency values presented in the gures have the following ND form Shh Shh , a0 GsB wB cs (5)

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546 547

From the gures, one can see that predictions of the present model for the foundation impedance values are in agreement with BEM values. This example is chosen to demonstrate the applicability of the present model for layered soils. Here, it should be noted that number of soil layers can be increased provided that each soil layer is dened by a nite region until the last layer extending to innity is reached. 4.4. Distributed transient strip load on an elastic halfspace The effect of a uniform 2B width strip transient load at the point A on the free soil surface which is 10B far away from the centre of the load on an elastic half-space is studied (see Fig. 11). The variation of the load with

548

549

550

551

552

553

554

555

520 521

556

557

522 523 524 525 526 971 972

558

559

560

561

973 975 974 977 976

Fig. 9.

Horizontal impedances of rigid strip foundation on a layer for GL / Gs

4.

4 5 2 1 3 2 1 9 380 981

ARTICLE IN PRESS
8 M.C. Genes, S. Kocak / Engineering Structures (2002)

982 984 983 986 985

Fig. 10. Horizontal impedances of rigid strip foundation on a layer for GL / Gs

0.25.

989 990

991 993 992 995 994

Fig. 11.

Distributed transient strip load on an elastic half-plane, geometry and loading.

997 998 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579

respect to time is given in Fig. 11. The constants dening the material behaviour are as follows: E 330MPa, n 0.25 and r 1.7857 103kg / m3. Due to symmetry, half of the problem is considered and the nite region is modelled by using 72 eight-node quadrilateral elements. The nite and innite region interface is modelled by using 18 three-node line elements. FFT algorithm is used for the transformation between frequency and time space. Since the frequency values at which innite soil responses are determined by the BulirschStoer algorithm automatically, interpolation is used for required values of Fourier parameters. Fig. 12 illustrates the results of the present model and Yerli et al. [7], which is a study based on using the innite elements on the truncated region to represent the innite soil medium. As can be seen in Fig. 12 the present model presents a good match with the model proposed by Yerli et al. [7]. 4.5. Rigid square foundation A massless rigid square foundation of width 2B on a layered half-space is considered. Similar to Example 4.2,

999 100

Fig. 12. Vertical displacement at point A.

100

580

581 582

to study how far the truncated soil region should be located, the size of the FE mesh given in Fig. 13(a) is decreased as in Example 4.2. The nal mesh is obtained by introducing a single layer of soil, as presented in Fig. 13(b). Due to symmetry, only a quarter of the problem is analysed, and a total of three meshes are considered

583

584

585

586

587

588

4 5 2 1 3 2 1 1 3006 1007

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 9

1008 1010 1009 1011 1012

Fig. 13. FE meshes (a) 27 elements for near-eld (b) nine elements for near-eld.

102 102 589 590 591 592 593 594 595

having 27, 18 and 9 twenty-node 3D solid nite elements for the nite region and 27, 21 and 15 eight-node nite elements for the interface. To demonstrate the applicability of the present model to various different problems, the following soil regions are considered: (i) homogeneous, (ii) layered, and (iii) top layer resting on a rigid bedrock. (i) Homogeneous soil. In this case, the top layer has a thickness of H 3B and the bottom layer is taken as half-space. The properties of the two layers are taken as identical with the following constants: Gs 24 109N / m2, ns 1 / 3, rs 2400kg / m3. The compliance values of the foundation calculated by the present model for various meshes are compared with others in Figs. 1416. In the diagrams the following ND compliance denitions are used Cvv GBCvv, Chh GBChh, Crr GB3Crr (6)

597 596 598 599 600 601 602 603 604 605 606

102 102

Fig. 15. Horizontal compliances of the square rigid foundation.

102

103 103

608 607

609 610 611 612 613 1014 1015

for vertical, horizontal and rocking compliances, respectively. For comparison, the available data are taken from different references. For vertical compliance, BEM based solutions are taken from Israil et al. [9] and innite element based solutions

103 103

Fig. 16.

Rocking compliances of the square rigid foundation.

103

1016 1018 1017 1020 1019

Fig. 14.

Vertical compliances of the square rigid foundation.

are taken from Yerli [31]. For horizontal compliance values, innite element based model of Yerli [31], and HallLamp and HallKitamura point loading formulation based model by Chow [32] are used. For rocking compliance, the results given by Yerli [31] are used. As in Example 4.2, the meshes studied presented close results (see

614

615

616

617

618

619

620

4 5 2 1 3 2 1 1 3038 1039

ARTICLE IN PRESS
10 M.C. Genes, S. Kocak / Engineering Structures (2002)

Svv

Svv , Shh GsB

Shh GsB

(7)

644 645

1040 1042 1041 1043 1046 1045 1044

Fig. 17. layer.

Vertical impedances of the square rigid foundation on a

621 622 623 624 625 626 627 628 629 630 632 631 633 634 635 636 637 638 639 640 641 642 643 1048 1049

Fig. 14), and for the sake of simplicity in Figs. 15 and 16, only nine element FE mesh results are plotted. From the Figs. 1416 one can see that the present model demonstrates a good match with the others. Similar to the example presented in Section 4.2, the advantage of the CIFEM for embedded structures can also be seen here in Fig. 14 by studying the good match between large mesh and the others. Here, large mesh represents the interface for embedded structures. (ii) Layered soil. In this case, top layer has a thickness of H B. The Poissons ratio for the two layers are taken as ns 1 / 3, and the ND properties of the two layers are taken as follows: GL 1, Gs 1.766, rL 1, rs 1.13 where subscripts L and s represent layer and half-space, respectively. In the analyses, the FE mesh given in Fig. 13(a) is used for modelling the soil layer and the impedance values of the foundation are compared with the results given by Tanrikulu [11] in Figs. 17 and 18. The ND vertical and horizontal impedances are given as follows:

In Figs. 17 and 18 the results predicted by the present model are found to be in agreement with the BEM results [11]. The BEM mesh used in [11] consists of 256 constant quadrilateral elements at the surface of the layer and 64 elements at the interface between the layer and half-space for the full problem. (iii) Top layer resting on a rigid bedrock. The top soil layer has the same properties with the one given in Case (i) and it has a thickness of H 2.5B. Here, to represent the rigid bedrock, the bottom soil layer is taken as rigid. This is accomplished by simply restraining the nodes on that surface. Here, to represent the soil layer, a nite region has to be dened until the rigid bedrock is reached. For this problem the FE mesh given in Fig. 13(a) is chosen, and in Fig.19 the ND vertical compliance values of the foundation are compared with the innite element results [31] and the BEM results obtained by computer program of Ref. [27]. In the diagram ND compliance is calculated as in Eq. (6). As can be seen from Fig. 19, the compliance values of the present model are in agreement with the others. 4.6. Parallel efciency of the model A parallel code for BulirschStoer algorithm is developed for applications on distributed systems. In order to investigate the efciency of the parallel algorithm, several test cases are studied. The studied cases showed that the parallel algorithm is efcient and easy to use. In this section, for the sake of simplicity, one example problem is given. Three different example meshes are chosen for structuremedium interface of example problem given in Section 4.5, and timings of

646

647

648

649

650

651

652

654 653

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

105 105

1050 1052 1051 1053 1056 1055 1054

106 106

Fig. 18. layer.

Horizontal impedances of the square rigid foundation on a

Fig. 19. Vertical compliances of the square rigid foundation on a layer resting on rigid bedrock.

106

106

4 5 2 1 3 2 1 3 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 11

those meshes are taken on WINDOWS and UNIX systems. The number of eight-node plane elements at the structuremedium interface are chosen as 27, 32 and 48 for meshes 1, 2 and 3, respectively. Here, it should be noted that the size of the dynamic-stiffness matrix to be calculated by BulirschStoer algorithm increases with the number of elements dened at the structuremedium interface. In a study by Kocak and Akay [25] the direct relation between the high performance parallel architectures and the efciency is presented. In that study, it is also shown that as the problem gets larger, higher speedups are achieved due to low communication costs. In this study, the time statistics demonstrated similar results to those of Kocak and Akay [25]. As UNIX based parallel system performed better than WINDOWS, for larger problems the efciency results noted better performance. In Fig. 20, the elapsed speedup diagrams of three meshes, pertaining to UNIX based system, are plotted. Here, it should be noted that the timings are not taken at reserved times, which means the processors of the system were also available to other users. In order to present more reliable results, the same problem is solved several times to reach to less-densely used system hours. As can be seen from the diagram, for the largest mesh, due to its being computation-bound, a maximum of ~6 speedup is reached for a twelve processor case and a decrease tendency starts after twelve processors. A similar observation is true for meshes 1 and 2 after eight processors where they reach to their maximum efciency. However, here it should be noted that for mesh 3, ~83% computation time is saved by using twelve processors. This is a signicant amount of time.

ent analyses FFT is used. Several example problems are solved to validate and investigate the applicability of the model. The results of the present model are compared with the data in the literature, and a good match is noted. By using the present model one can introduce a small computational mesh for innite soil region compared to other Finite or Boundary Element based models. CIFEM requires the solution of a rst-order non-linear ODE for the calculation of dynamic-stiffness matrix of the innite soil region. This differs from other FE or BE based methods where the dynamic-stiffness matrices can be calculated in discrete frequency values by simply introducing inertia forces. This results in the use of an integration algorithm, which is highly expensive in computation time. As a remedy to this problem, the integration algorithm is parallelized for applications on concurrent systems. Especially for large-scale problems, signicant computational time is saved, e.g. for a moderately large problem 83% of total elapsed time is saved by using twelve processors.

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

Acknowledgements The authors are thankful to Dr Y. Mengi and Dr A.H. Tanrikulu for their permission to use their BEM programs for comparisons.

738

739

740

741

Appendix A By using the similarity [1], another similar ctitious interface is dened by multiplying the coordinates of the interface between near- and far-eld according to similarity centre O by a similarity factor (1 w). The similarity factor is dened by dividing the distance between the similarity centre and the exterior (ctitious) boundary by the distance between the similarity centre and interior boundary. In the CIFEM, the ND cell width measured in the radial direction is assumed to be a very small number (see Fig. 1). Therefore, the relation between the dynamic-stiffness matrices of the interior and exterior boundaries will yield a limit with ND cell width (w). This results in a derivative of the dynamic-stiffness matrix of the unbounded medium with the characteristic length r of the interior boundary according to similarity centre. This derivative of the dynamic-stiffness matrix according to r can be altered with the derivative of the dynamic-stiffness matrix according to frequency w based on similarity in frequency domain [1] as given in Eq. (A.1) rS (w),r (s 2)S (w) wS (w),w (A.1)

742

743

744

745

713

5. Conclusions A combined FE based SSI model is studied. In the model, for innite soil region CIFEM in frequency space is used, and nite soil region is added on top. The model is subjected to harmonic and transient loads. For transi-

746

747

714 715 716 717

748

749

750

751

752

1067 1068

753

754

755

756

757

758

759

760

761

762

763

765 764

1069 1071 1070 1073 1072

Fig. 20. Speedup diagram for BulirschStoer algorithm.

where, S (w) is the dynamic-stiffness matrix of the unbounded medium and s is the spatial dimension (=2

766

767

4 5 2 1 3 2 1 3 768 769 770 771

ARTICLE IN PRESS
12 M.C. Genes, S. Kocak / Engineering Structures (2002)

or 3). This represents the dynamic-stiffness matrix of the unbounded medium as a function of frequency w, and results in a system of non-linear ODE of rst-order in the independent variable w (S (w) E ) (E ) (S (w)
2 1 0 1

772

(E ) ) (s wM
2 0

1 T

(A.2) 0

774 773

2)S (w) wS (w),w E

775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797

where, E0, E1, E2 represent the static-stiffness matrices and M0 represents the mass matrix of the interior surface of the cell [1]. The solution of the system of non-linear ODE given in Eq. (A.2) can be performed by using BulirschStoer method [22,23], which is a powerful and efcient integration algorithm. For the solution of the differential equation by BulirschStoer method, the initial starting dynamic-stiffness matrix of the interface must be known. In order to propose an acceptable solution, the boundary condition must satisfy the radiation condition in frequency domain, and yet the static-stiffness matrix cannot satisfy the radiation condition explicitly. The radiation condition must be formulated at an innite distance from the interface, which corresponds to an innite characteristic length r. However, by introducing the associated ND frequency a0 wr / cs, innite ND frequency can also be achieved by an innite w instead of an innite r. In an actual calculation, the dynamic-stiffness matrix to be used in Eq.(A.2) can be calculated by choosing a large w. Thus, the high-frequency behaviour of S is studied examining the asymptotic expansion of the differential equation given in Eq. (A.3)
m

798 799

S (w) iwC

K
j

1 A (iw)j j 1

(A.3)

800 801 802 803 804 805 806 807

where C , K and Aj represent dashpot matrix, spring matrix and unknown coefcient matrices of the asymptotic expansion, respectively. m in Eq.(A.3) represents the number of terms to be considered in the analysis, and i is the imaginary number. The details of the calculation of S can be found in Ref. [1]. Once S at a high a0 (ND frequency) is calculated, it remains to solve Eq. (A.2) by BulirschStoer integration. References
[1] Wolf JP, Song C. Finite-element modelling of unbounded media. England: Wiley, 1996. [2] Lysmer J, Kuhlemeyer RL. Finite dynamic model for innite media. J Eng Mech ASCE 1969;95(EM4):85977. [3] White W, Valliappan S, Lee IK. Unied boundary for nite dynamic models. J Eng Mech ASCE 1977;103(5):94964. [4] Bettess P. Innite elements. Int J Numer Methods Eng 1977;11:5364. [5] Medina F, Penzien J. Innite elements for elastodynamics. Earthquake Eng Struct Dynamics 1982;10:699709. [6] Yang S, Yun C. Axisymmetric innite elements for soilstructure interaction analysis. Eng Struct 1992;14(6):36170.

808

809 810 811 812 813 814 815 816 817 818 819 820

[7] Yerli HR, Temel B, Kiral E. Transient innite elements for 2-D soilstructure-interaction analysis. J Geotech Geoenviron Eng Div ASCE 1998;124(10):97688. [8] Karabalis DL, Beskos DE. Dynamic response of 3-D rigid surface foundations by time domain boundary element method. Earthquake Eng Struct Dynamics 1984;12:7393. [9] Israil AS, Banerjee PK. Effect of geometrical and material properties on the vertical vibration of three-dimensional foundations by BEM. Int J Numer Anal Methods Geomech 1990;14:4970. [10] Von Estorff O, Pais AL, Kausel E. Some observations on time domain and frequency domain boundary elements. Int J Numer Methods Eng 1990;29:785800. [11] Tanrikulu AH, Yerli HR, Tanrikulu AK. Application of the multiregion boundary element method to dynamic soilstructure interaction analysis. Comput Geotech 2001;28:289307. [12] Von Estorff O, Kausel E. Coupling of boundary and nite elements for soilstructure interaction problems. Earthquake Eng Struct Dynamics 1989;18:106575. [13] Von Estorff O, Firuziaan M. Coupled BEM/FEM approach for non-linear soilstructure interaction. Eng Anal Boundary Elements 2000;24:71525. [14] Yu G, Mansur WJ, Carrer JAM, Lie ST. A more stable scheme for BEM/FEM coupling applied to two-dimensional elastodynamics. Comput Struct 2001;79:81123. [15] Wolf JP, Song C. Dynamic-stiffness matrix in the time domain of unbounded medium by innitesimal nite-element cell method. Earthquake Eng Struct Dynamics 1994;23:118198. [16] Wolf JP, Song C. Unit-impulse response matrix in time domain of unbounded medium by innitesimal nite-element cell method. Comput Methods Appl Mech Eng 1995;122:25172. [17] Song C, Wolf JP. The scaled boundary nite-element methodalias consistent innitesimal nite-element cell method for elastodynamics. Comput Methods Appl Mech Eng 1997;147:32955. [18] Wolf JP, Song C. The scaled boundary nite-element method-a primer: derivations. Comput Struct 2000;78:191210. [19] Song C, Wolf JP. The scaled boundary nite-element method a primer: solution procedures. Comput Struct 2000;78:21125. [20] Cooley JW, Lewis PAW, Welch PD. The Fast Fourier transform and its applications. IEEE Trans Educ 1969;12:2734. [21] Bathe KJ. Finite element procedures in engineering analysis. New Jersey: Prentice-Hall, 1982. [22] Flannery WH, Teukolsky BP, Vetterling WT. Numerical recipes. Cambridge: Cambridge University Press, 1988. [23] Bulirsch R, Stoer J. Numerical treatment of ordinary differential equations by extrapolation methods. Numerische Mathematik 1966;8:113. [24] Farhat C, Wilson E, Powell G. Solution of nite element systems on concurrent processing computers. Eng Comput 1987;2:157 65. [25] Kocak S, Akay HU. Parallel Schur complement method for largescale systems on distributed memory computers. Appl Math Modelling 2001;25:87386. [26] Genes MC. Soil structure interaction models for 2-D and 3-D problems and applications on parallel platforms. PhD Dissertation, Cukurova University, Department of Civil Engineering; 2001. [27] Mengi Y, Tanrikulu AH, Tanrikulu AK. Boundary element method for elastic media an introduction. Ankara: Middle East Technical University Press, 1994. [28] Luco JE, Westman RA. Dynamic response of a rigid footing bounded to an elastic half-space. J Appl Mech ASME 1972;39(2):52734. [29] Temel B. Dynamic analysis of soilstructure interaction problems by using nite-innite elements and Laplace transformation method. PhD Dissertation, Cukurova University, Department of Civil Engineering; 1996.

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

4 5 2 1 3 2 1 3 888 889 890 891 892 893

ARTICLE IN PRESS
M.C. Genes, S. Kocak / Engineering Structures (2002) 13

[30] Tanrikulu AH. A boundary element model with nonlocal boundary conditions for dynamic analysis of a two-phase composite. PhD Dissertation, Cukurova University, Department of Civil Engineering; 1999. [31] Yerli HR. 2D and 3D analyses of dynamic soilstructure interaction problems by coupling of nite and innite element. PhD

Dissertation, Cukurova University, Department of Civil Engineering; 1998. [32] Chow YK. Simplied analysis of dynamic response of rigid foundations with arbitrary geometries. Earthquake Eng Struct Dynamics 1986;14:64353.

894

895

896

897

900 899 898