FatiguePredictions.pdf

Prediction of the Fatigue Life of Cast Steel ContainingShrinkage Porosity

RICHARD A. HARDIN and CHRISTOPH BECKERMANN

A simulation methodology for predicting the fatigue life of cast steel components with shrinkageporosity is developed and validated through comparison with previously performed measure-ments. A X-ray tomography technique is used to reconstruct the porosity distribution in 25 testspecimens with average porosities ranging from 8 to 21 pct. The porosity eld is imported intonite element analysis (FEA) software to determine the complex stress eld resulting from theporosity. In the stress simulation, the elastic mechanical properties are made a function of thelocal porosity volume fraction. A multiaxial strain-life simulation is then performed to deter-mine the fatigue life. An adaptive subgrid model is developed to reduce the dependence of thefatigue life predictions on the numerical mesh chosen and to account for the eects of porositythat is too small to be resolved in the simulations. The subgrid model employs a spatiallyvariable fatigue notch factor that is dependent on the local pore radius relative to the niteelement node spacing. A probabilistic pore size distribution model is used to estimate the radiusof the largest pore as a function of the local pore volume fraction. It is found that, with theadaptive subgrid model and the addition of a uniform background microporosity eld with amaximum pore radius of 100 lm, the measured and predicted fatigue lives for nearly all 25 testspecimens fall within one decade. Because the fatigue lives of the specimens vary by more thanfour orders of magnitude for the same nominal stress amplitude and for similar average porosityfractions, the results demonstrate the importance of taking into account in the simulations thedistribution of the porosity in the specimens.

DOI: 10.1007/s11661-008-9755-3 The Minerals, Metals & Materials Society and ASM International 2009

I. INTRODUCTION

INHOMOGENEITIES due to porosity are currentlynot considered in the design of structural componentsmade from metal castings. Instead, ad hoc safety factorsare used to address a designers uncertainty in how thecasting will perform in service. These safety factors arebased on the assumption that castings perform unpre-dictably, if not poorly. Applying such safety factors tothe entire cast material might do little for the robustnessof the design other than increase the casting weight.Many part designers become frustrated by castingsdesigned with very large safety factors that fail inservice; they are hesitant to use castings. Such frustra-tions could be avoided if the quality of the cast metalthroughout the casting could be known ahead of timeand incorporated into the design.The present study extends our recently developed

method of modeling the eects of porosity on stinessand stress redistribution[1] to the prediction of thefatigue life of steel castings containing porosity. Forease of use in standard design practice, commonly usedcommercial software is employed in the present stress

and fatigue life simulations. The part design can bemade safe by assuring that the casting process results inthe best possible quality (i.e., lowest porosity) steel athighly stressed locations. Should porosity form, thecasting rigging or process parameters can be changed sothat the porosity will not aect the service performanceof the part. If designers wish to use lighter-weight andmore thinly walled steel castings, understanding theeects of porosity becomes especially critical. Anintegrated design process is emerging in which a castingprocess simulation that predicts the location, amount,and size of the porosity is directly coupled with themechanical simulation of the part performance thattakes into account the eects of porosity.[2] It isanticipated that such a design process will also helpguide and improve casting inspection procedures, bylinking acceptance criteria with expected performance.Fatigue life analysis can be divided into two parts:[3]

(1) the stage of life of the component up to the initiationof a crack on the order of 1 mm in size and (2) the stageof life of the component undergoing the growth of acrack and its propagation to failure. The combination ofthe two gives the total life of a component. Perhaps themost often used approach to predicting crack initiationis the strain-life method,[36] in which specimen fatiguelife test data to failure is tted to the strain-life curve andthe curve is used in life estimation. Because fatigue testspecimens are small compared to most components,once cracks initiate in them, propagation is very rapid.The total life from the strain-life testing of specimens is

RICHARD A. HARDIN, Research Engineer, and CHRISTOPHBECKERMANN, Professor, are with the Department of Mechanicaland Industrial Engineering, University of Iowa, Iowa City, IA 52242.Contact e-mail: [email protected]

Manuscript submitted January 30, 2008.Article published online January 21, 2009

METALLURGICAL AND MATERIALS TRANSACTIONS A VOLUME 40A, MARCH 2009581

the crack initiation life, because the second stage of lifeis extremely short relative to the rst, for these testconditions. Multiaxial strain-life models[712] have beenimplemented in commercial software packages (fe-safe,for example[13]), to predict the durability of componentsfrom nite element analyses (FEAs). The second stageof fatigue life, the fatigue crack growth stage, isdescribed using fracture mechanics, in particular, linearelastic fracture mechanics (LEFM), in which crackgrowth is assumed to follow the ParisErdogan equa-tion.[14] Here, a crack pre-exists, formed either fromfatigue crack initiation or during manufacturing of thepart; the crack growth is determined by the stress eldand material properties until failure.[3,14] Because thelargest possible nonpropagating crack can be predictedfor a given material and an alternating stress eld usingfracture mechanics, it can be used in damage-tolerantdesign life prediction in castings.[1517]

The eects of porosity on the fatigue behavior of castmetals have been measured in steels,[1824] in castirons,[16,17,25,26] and in aluminum alloys.[2736] Also,Murakami[37] gives a good overview of the eects ofsmall defects and inclusions on metal fatigue. Jayet-Gendrot et al.[20] have reviewed the fatigue and fractureliterature for steel and other metals and showed that thefatigue behavior depends on the pore size and volumefraction. For example, in stainless steel, they report thatreductions in fatigue strength of 35 and 50 pct areobserved for areas of crack initiating defects of less thanand greater than 3 mm2, respectively, in a specimen testsection 70 9 22 mm in size.[21] Based on the square rootof the defect area as a representation of the dimension ofthe porosity,[37] this size is approximately 55 lm. Also,for low-alloy steel, fatigue strength reductions from 8 to30 pct were reported when shrinkage porosity cavitiescovered 3 to 7 pct of the fracture surface.[22] Recently,for an aluminum alloy, Linder et al.[29] measured a15 pct decrease in fatigue strength for smooth specimenswhen the porosity level was increased from 0.7 to4.1 pct, while notched specimens gave no dependenceon porosity in the same range. Extreme valuestatistics[2830,34,37] have been used to show the strongdependency of fatigue life on the maximum likely poresize from measured pore size distributions.Crack initiation life analyses of cast components in

the presence of porosity have been performed by usingthe strain-life approach and modeling the pores asequivalent notches.[1824,26,3133] Applying a strain-lifemodel alone assumes that crack nucleation encompassesthe majority of the life of the component and that thetime for fatigue crack propagation to fracture isinsignicant relative to crack initiation. The use ofstrain-life models requires that pore geometry informa-tion, primarily the minimum notch radius and the majoraxes of the ellipsoidal notch, be known or be determinedfrom fracture surfaces, to determine a stress concentra-tion factor. Researchers studying aluminum castingshave treated pores in castings as notches for predictingthe eect of porosity on crack initiation life[31] and themechanism of crack formation and growth frompores.[33] For the second stage of component life,fatigue life estimation of cast metals by fracture

mechanics approaches are used: for nodular castiron,[16,17,25,26] cast aluminum alloys,[27,3236,38] and steelalloys.[15,19,20,23] One diculty in applying fracturemechanics concepts to cast parts is determining thenal, or failure, crack length. For example, failure ofcomponents has been assumed when the remaining netsection area stress is at or is greater than the yieldstrength[19] and when the crack depth propagates to thesection wall thickness at the defect location.[15] Acomparison between the measured fatigue life of caststeel test specimens containing porosity and inclusionsand the fatigue life obtained by modeling the specimenusing crack initiation (local stain-life concepts) andLEFM approaches has been presented.[23] In the crackinitiation model, the defects were considered to be three-dimensional notches; in the crack growth model, thedefects were treated as two-dimensional elliptical crackshaving an envelope around the defect. It was found[23]

that the crack initiation estimate of life was moreaccurate than the fracture mechanics approach, andthat interpreting the porosity as pre-existing cracksresulted in too conservative an estimate of the fatiguelife. Dabayeh and Topper[32] came to a somewhatdierent conclusion for cast aluminum, in which thelocal strain approach gave quite nonconservative esti-mates of fatigue life and the crack growth method gavebetter agreement. The current authors have found thatlocal strain-life modeling of porosity as a notch gavebetter agreement with measurements than did LEFM incast 8630 steel.[19]

In the present study, a simulation method is devel-oped to predict the measured fatigue lives of 25 cast steelspecimens containing up to 21 pct shrinkage porosityover the gage length. This is a challenging task, not onlybecause the measured fatigue lives of the specimens aremuch below the values corresponding to sound steel, butalso because the measured lives vary by more than fourorders of magnitude at the same applied stress ampli-tude.[19] These variations can only be attributed to thedierent amounts and distributions of porosity withinthe specimens. The fatigue tests and the tomographicreconstruction of the porosity elds in the specimens[1]

are briey reviewed in Section II. In Section III, theprocedures and results of the fatigue life simulations arepresented. An FEA is performed to determine thecomplex stress eld resulting from the porosity;[1]

multiaxial strain-life analysis is used to compute thefatigue life distribution in the specimens. The strain-lifeapproach is adopted here, because the preliminarycalculations in Reference 19 revealed that LEFM doesnot predict the measured fatigue lives of the specimens.An adaptive subgrid fatigue notch factor model, whichapproximately accounts for the eect of porosity that isunder-resolved by the computational mesh used in thestress and fatigue life simulations, is presented in SectionIV. As part of this subgrid model, a probabilistic poresize distribution model for determining the maximumpore radius as a function of the local pore volumefraction is developed. Detailed comparisons are madebetween the measured and predicted fatigue lives. Theconclusions of the present study are summarized inSection V.

582VOLUME 40A, MARCH 2009 METALLURGICAL AND MATERIALS TRANSACTIONS A

II. EXPERIMENTS

Fatigue tests were performed using cast and heattreated AISI 8630 steel specimens containing a range ofshrinkage porosity. All the details of these experiments,together with the results of fractography and prelimi-nary fatigue life calculations, were presented in Refer-ence 19. Prior to the fatigue testing, X-ray tomographywas performed on the machined specimens, in order toobtain a measured three-dimensional porosity eldwithin each specimen. The details of the tomographicreconstruction procedures were recently presented inReference 1. For completeness, a brief overview of thesetests and measurements is provided here.

A. Cast Specimens

Specimen blanks were cast from AISI 8630 steel. Themold geometry was designed using computer modeling,to obtain a range of shrinkage porosity levels.[19] Thecast blanks were 152-mm-long cylinders having anominal 14.3 mm diameter. To produce shrinkageporosity, a cylindrical disk 25.5 mm in diameter waspositioned at the midlength of the blanks, as shown inFigure 1. This design concentrated the porosity at thecenterline and midlength of the cast blanks, so that theporosity could be located in the gage section of the testspecimens. The severity of the porosity was controlledby varying the disk thickness (dimension along thecasting length); disk thicknesses of 5, 7.5, and 10 mmwere cast. Generally, a smaller disk thickness resulted in

a lower porosity level, but there was overlap ofthe porosity levels among the three disk thicknessgroups. Cut surfaces of specimens from castings havingdisk thicknesses of 5, 7.5, and 10 mm are shown inFigures 2(a), (b), and (c), respectively. It can be seenthat the porosity level ranges from dispersed macropo-rosity to holes and gross section loss. Radiographs areshown to the right of each cut specimen surface with a

X 152.0

14.25 25.5

Fig. 1Dimensions of cast porous specimen blanks in millimeters.Dimension X = 5, 7.5, and 10 mm for the least, middle, andmost porosity specimen groups, respectively.

Fig. 2Cut and polished surfaces of three specimens cast with dif-ferent porosity levels ranging from (a) specimen 22 representing theleast, to (b) specimen 3 representing the middle range, to (c) speci-men 13 representing the most porosity. Radiographs of the specimengage sections are given to the right of each surface, with the longitu-dinal position of the cut indicated.


white mark that indicates the approximate location ofthe cut along the length of the machined specimen.Comparing these radiographs qualitatively, the porosityappears to be increasingly severe proceeding fromFigure 2(a) to Figure 2(b), and to Figure 2(c). In somespecimens, as shown, for example, in Figure 2(c),porosity extended to the specimen surface.The cast blanks received identical heat treatment: they

were normalized at 900 C, austenized at 885 C, waterquenched, and nally tempered for 1 1/2 hours at510 C. This heat treatment resulted in a temperedmartensitic structure with a Rockwell C hardness of 34.The blanks were then machined to the dimensionsshown in Figure 3. More than 50 specimens wereproduced in this manner.[19]

B. Tomography for Porosity Measurement

The X-ray tomography of the machined specimenswas performed in order to obtain a measured three-dimensional porosity eld within each specimen. Thismethod is described in detail in Reference 1. In brief,lm radiography of the specimens was performed usinga sensitivity of 2 pct of the gage section diameter, whichcorresponds to a resolution of approximately 100 lm.Three examples of radiographs are shown in Figure 2.Orthogonal radiographic views of each specimen wereobtained. The radiographic lms were digitized usingan X-ray scanner resulting in 8 bit gray level, 1200 dpi(~21 lm pixel side length) images. The gray levelinformation was converted to porosity percentage usingan in-situ calibration method. A tomography algo-rithm[1,39] was then employed to reconstruct a three-dimensional porosity eld from the orthogonal views.Figure 4 shows two examples of porosity elds

obtained in this manner. The X-ray tomography tech-nique used cannot resolve the exact geometry of theshrinkage pores on a microscopic scale. This wouldrequire a resolution on the order of 1 lm over aspecimen gage length and diameter of 17.5 and 5 mm,respectively, which is not possible given present daycomputing resources. The porosity volume fractionsobtained in the tomographic reconstruction correspondto voxels that have a side length of ~21 lm. If a voxelfalls entirely within a pore (or sound steel), the porosityvolume fraction is equal to unity (or zero). In thesimulations presented here, the measured porosity eldsare mapped onto computational meshes with nodespacings that are one to two orders of magnitude larger

than the voxel side length, i.e., between 100 lm and1 mm. Thus, the accuracy and resolution of the presenttomographic reconstruction technique were deemed tobe sucient (also Reference 1).Table I lists the average porosity volume fraction

measured for the 25 specimens for which a tomographicreconstruction was performed. The average porosityfractions listed in the table correspond to a 12 mm longcenter portion of the gage section, rather than the entire17.5 mm gage length (Figure 3) over which the porosityeld was reconstructed; this is related to the location ofthe extensometer with which the strain was measured(Section C). It can be seen that the average porosityvaries from approximately 8 to 21 pct. Table I also liststhe maximum cross-sectional porosity fraction for eachspecimen. This fraction was obtained by averaging theporosity over one voxel thick cross-sectional slices andselecting the maximum value along the gage length. Itvaries from approximately 15 to 59 pct. It is shown inReference 1 that this maximum cross-sectional porosity,rather than the average volume fraction over the gagelength, correlates well with the measured eective elasticmodulus, E, of the specimens.

C. Fatigue Testing

The specimen preparation and fatigue testing werecarried out according to the ASTM E606 standard.[4] Allfatigue tests were performed under fully reversed,R = 1, loading conditions. During fatigue testing,the mechanical behavior of the specimens was found tobe elastic; therefore, load-controlled testing at 10 to20 Hz was used. This permitted accurate strain ampli-tude measurements while using the faster testing capa-bility of load control. All fatigue tests were performeduntil fracture of the specimen occurred or a runout lifewas achieved at 5 9 106 cycles. Four nominal stressamplitudes, DS/2, were applied and held constant duringtesting of the specimens: 126, 96, 66, and 53 MPa; thesestress levels were selected to obtain a large range ofmeasured fatigue lives while avoiding runouts. Anextensometer with ends at 6 mm above and below themidpoint of the specimen length was used to measurethe strain amplitude. The eective (or apparent) elasticmodulus E was determined from the measured strainand load data, using the sound specimen gage cross-sectional area to determine the nominal stress. Themeasured elastic moduli and fatigue lives of the 25specimens for which the measured fatigue live was lessthan the runout life are given in Table I; reference willbe made to individual specimens in the table using thespecimen numbers.The monotonic and cyclic material properties for

sound 8630 steel were measured in Reference 40 and areprovided in Tables II and III, respectively. The cyclicstress-strain curve is given by

De2 DS2E

DS2K0

1n0

1

where DS/2 and De/2 are the nominal stress and strainamplitudes, respectively; K is the cyclic strength

116.5

38.5 11 17.5

5 .025

12R 19

Gage section

Fig. 3Dimensions of test specimens in millimeters.


coecient; and n is the cyclic strain hardening expo-nent. Note from Tables II and III that the 661 MPavalue for the cyclic yield strength, S0y, is much less

than the monotonic yield strength, Sy, value of985 MPa, indicating the presence of considerablecyclic softening. The strain-life curve (often referred to

Fig. 4Example of internal porosity distributions reconstructed from X-ray tomography in the test section of fatigue test specimens 8 and 22.Slices at three axial positions are shown for each specimen.


as the ConManson relationship[5,6]) for sound 8630steel is given by

De2 Dee

2 Dep

2 r

0f

E2Nf b e0f 2Nf c 2

where De/2, Dee/2, and Dep/2 are the total, elastic, andplastic strain amplitudes, respectively; 2Nf is the numberof reversals to failure; r0f is the fatigue strength coe-cient; b is the fatigue strength exponent; e0f is the fatigueductility coecient; and c is the fatigue ductilityexponent. The sound 8630 steel cyclic properties listedin Table III were obtained in Reference 40 by tting testdata to Eqs. [1] and [2]. For sound 8630 steel, the fatiguestrength (or fatigue limit), Sf, at 5 9 10

6 cycles is equalto 293 MPa.The measured fatigue lives (as cycles to failure, Nf) of

the porous specimens are plotted in Figure 5 as afunction of the applied stress amplitude, DS/2. InFigure 5(a), the measurements are compared to thestress-life curve for sound 8630 steel. The sound stress-life curve was obtained from Eq. [2] by converting thestrain to stress using the elastic modulus from Table II;plasticity was found to be negligibly small.[19] It can beseen that the fatigue lives of the porous specimens fallfar below the sound material curve. Depending on thenumber of cycles to failure, the stress amplitudes in thepresent tests are between approximately 170 and900 MPa below the stress amplitudes for sound 8630

Table I. Summary of Measurements for 25 Cast Steel Specimens Containing Porosity:[1,19] Test Stress Level, Porosity fromRadiographic Analysis, Strain, Elastic Modulus, and Fatigue Life

SpecimenNumber

AppliedStress Levelof Test(MPa)

AveragePorosity

Volume Fractionin Gage Section

MaximumCross-Sectional

PorosityFraction Strain 9104

ElasticModulus(GPa)

FatigueLife

(Cycles)

1 96 0.104 0.432 7.01 137 13652 96 0.101 0.208 6.44 149 79,9083 126 0.101 0.256 8.81 143 24,3204 53 0.128 0.297 3.84 138 851,2755 126 0.097 0.185 8.24 153 29,0236 66 0.134 0.275 4.55 145 216,5167 66 0.095 0.201 4.68 141 4,053,8008 66 0.185 0.363 4.89 135 57,5669 53 0.146 0.587 6.09 87 10,81210 126 0.213 0.326 9.33 135 37,08911 66 0.117 0.250 4.85 136 113,50312 66 0.148 0.415 5.84 113 15,41913 96 0.185 0.538 12.47 77 604214 126 0.139 0.507 10.50 120 16015 53 0.187 0.551 6.09 87 15,86816 66 0.076 0.160 3.98 166 1,681,01817 96 0.096 0.254 8.65 111 439218 53 0.085 0.148 3.66 145 1,342,21819 126 0.093 0.225 8.87 142 13,01320 53 0.099 0.182 3.71 143 249,75221 96 0.122 0.300 7.68 125 41,06622 66 0.117 0.262 4.37 151 769,07423 126 0.121 0.205 8.13 155 40,89624 96 0.144 0.288 6.76 142 333,02525 126 0.129 0.280 8.51 148 7456

Table II. 8630 Steel Monotonic Properties[40]

Property Sound Material

Su (MPa) 1,144Sy (MPa) 985E0 (GPa) 207Pct EL not measuredPct RA 29rf (MPa) 1,268ef 0.35K (MPa) not measuredN not measured

Note: EL = elongation, RA = reduction in area.

Table III. 8630 Steel Cyclic Properties[40]

Property Sound Material

Sf (MPa) 293Sf/Su 0.26K (MPa) 2,267n 0.195S0y (MPa) 661b 0.121c 0.693r0f (MPa) 1,936e0f 0.42


steel. If the specimens had been sound, they would allhave a life greater than the runout value (>5 9 106

cycles). The porous specimen data is examined in moredetail in Figure 5(b). For all applied stress levels, largevariations in the measured fatigue lives can be observed.The lives of the porous specimens range from 160 cyclesto a runout. Clearly, the fatigue behavior of thesespecimens is primarily controlled by porosity. However,the variation in the measured fatigue life of more thanfour orders of magnitude cannot be explained by thecomparably small dierences in the average porosity ofthe specimens (between 8 and 21 pct (Table I)). Theseresults demonstrate that each specimen has a uniqueporosity eld; this, in combination with the loading,results in a complex stress eld that must be accuratelysimulated in order to predict the fatigue life of a givenspecimen.

III. FATIGUE LIFE PREDICTION

A. Simulation Procedure

The present procedure for predicting the fatigue life ofthe porous specimens can be summarized as follows.

(a) Map the measured porosity eld from the tomo-graphy onto the nodes of the FEA mesh.

(b) Degrade the elastic properties at each node accord-ing to the local porosity fraction, /.

(c) Perform a nite element elastic stress analysis thatcorresponds to the loading in the fatigue tests.

(d) Import the predicted stress elds correspondingto tension and compression into the life predic-tion software and perform a multiaxial strain-lifeanalysis.

The rst three steps in this procedure were developedpreviously by the present authors; all details can befound in Reference 1. In brief, the three-dimensionalquadratic interpolation subroutine QD3VL from theInternational Mathematics and Statistics Library(IMSL)[41] is used to map the porosity data onto theFEA mesh. Ten-node quadratic tetrahedral elements areused to perform the stress analysis. Elastic mechanicalproperties are assigned at each node as a function of theporosity fraction at that location. The local elasticmodulus is calculated from[1]

E / E0 1 /0:5

2:53

where E0 is the elastic modulus of the sound materialand / is the porosity volume fraction. The Poissonratio, m, as a function of / is obtained from a relation-ship developed by Roberts and Garboczi:[42]

m / mS //1m1 mS 4

with m = 0.14, / = 0.472, and mS = 0.3. Signicantplasticity was not detected during testing of the speci-mens; therefore, plastic eects are ignored in the FEAsimulations.[19] Mechanical simulation of the specimenswith porosity is performed using ten-node quadratictetrahedral stress/displacement elements (type C3D10)with the commercial FEA package Abaqus/Standard(version 6.6.1).[43] The FEA simulation boundary con-ditions are chosen to closely match the test conditions.During testing, the specimens were held xed at theirupper grip, and the loading was applied to the lower(ram) end, which was free to move vertically, placing thespecimen in tension and compression during fatiguetesting. The simulated specimen geometry considersonly the initial 5 mm of the length of the grips in theFEA mesh. This is a distance away from the llet andtest section that is more than sucient for producingstress-strain results that are insensitive to the locationsat which the boundary and loading conditions areapplied. Shorter execution times are the only dierencefrom simulations using more of the grip length. At theupper grip face, a clamped boundary condition isapplied (having no translations or rotations); at thelower grip, translations are allowed only in the axialdirection, with no axial rotation allowed. A uniformdistributed loading is applied over the face at the lowergrip end, to produce the total load corresponding to thetesting conditions. Using this procedure, it is shown inReference 1 that the measured strains (Table I) arepredicted to within approximately 10 pct. This goodagreement is achieved for an FEA node spacing of

Runout at 5106 cycles

Fatigue Life, Nf (Cycles to Failure)

Runout at 5106 cycles

Fatigue Life, Nf (Cycles to Failure) 101 102 103 104 105 106 107

101 102 103 104 105 106 107

900Sound keel block data [19]Test Data

800

700

600

500

400

300

200

100

0

0

20

40

60

Stre

ss A

mpl

itude

(MPa

)St

ress

Am

plitu

de (M

Pa)

80

100

120

140

160

180

200

(a)

(b)

Fig. 5(a) Fatigue life measurements of specimens with porositycompared with sound data from Ref. 40. (b) Measured fatigue livesof specimens with porosity tested at four stress levels: 126, 96, 66,and 53 MPa. Note the single runout specimen.


0.25 mm, but the results are relatively independent ofthe mesh for node spacings up to 2 mm.[1] These resultsshow that, by using porosity fractions that are denedover a volume that is small compared to the (macro-scopic) specimen geometry but large compared to the(microscopic) pore geometry, and by degrading theelastic properties locally in accordance with the porosityfraction, the overall stiness response of the specimenscan be simulated accurately.After running ABAQUS, the resulting stress elds

corresponding to the tension and compression steps ofthe fully reversed loading (R = 1) are imported intothe fatigue life prediction software. In the present study,the software package fe-safe[13] is used for the fatigue lifecalculations. The loading cycle and the fatigue proper-ties for the sound material (Table III) are entered intothe fe-safe model. The nodal stress tensors from theAbaqus simulations are converted to strains within thefe-safe software, using a sound material elastic modulusof 207 GPa (Table II). As long as this elastic modulus isused throughout the fatigue calculations, the use of avariable E is not necessary as the alternating strainamplitude and the elastic term in the strain-life equationscale with whatever modulus is used. In the presentstudy, fe-safes multiaxial BrownMiller algorithm withthe Morrow mean stress correction is used to calculatethe fatigue life as cycles to failure. For steel, fe-saferecommends using the multiaxial BrownMiller algo-rithm for ductile steel and the principle strain algorithmfor brittle steels, both with the Morrow mean stresscorrection.[13] The BrownMiller algorithm is said to bea more conservative method; it uses a critical planeanalysis to determine the life in reversals to failure, 2Nf,by solving

Dcmax2

Den2

1:65 r0f rmE

2Nf b 1:75e0f 2Nf c 5

at each node, where Dcmax/2 is the maximum shearstrain amplitude, Den/2 is the strain amplitude normalto the shear stress plane, and rm is the mean stress. Thecritical plane is dened as the plane having themaximum value of Dcmax/2+Den/2. In the criticalplane analysis, the calculated strain tensor at a niteelement node (having three direct and three shearcomponents) is resolved onto a number of planes wherethe damage associated with the strain is evaluated oneach plane. The plane with the most damage is thenselected for use in the strain-life calculations. In aCartesian x-y-z coordinate system, unique planes can bedened by the orientation the normal of the planesurface makes with respect to the coordinate system.This orientation can be dened by one angle from thex-axis toward the y-axis, and a second angle from thez-axis toward the x-y plane.[13] The software fe-safesearches for the critical plane with the worst damage(shortest life) in 10 deg increments over the 180 degrange of the rst angle and the 90 deg range of thesecond angle. Direction cosines are used to project thestrains onto the calculation plane. Additional detailsabout multiaxial fatigue analysis can be found in thebook by Socie and Marquis.[7]

B. Results of Fatigue Life Simulations Using MultiaxialStrain-Life Method

Figure 6 shows a series of simulation results forspecimen 19 on a longitudinal center section, startingwith the measured porosity eld mapped onto the FEAmesh and followed by the predicted maximum principlestress and strain elds and, nally, the predicted cyclesto failure eld. The latter is also shown on the surface ofthe specimen. In these baseline simulations, a nodespacing of 0.25 mm is used. As mentioned previously,this grid of approximately 20 nodes across the specimendiameter was found to give good results for the overallstiness.[1] It can be seen that the porosity causes acomplex three-dimensional stress eld to develop in thespecimen during loading. Larger stresses are observedadjacent to regions with a high porosity fraction. On theother hand, large strains coincide with high porosityfractions. The shortest lives are predicted in regions ofhigh stress concentration. For this specimen, the node

Fig. 6Results in a longitudinal section for specimen 19. (a) Poros-ity distribution from tomography. Results from Abaqus stress analy-sis for (b) maximum principal stress and (c) maximum principalstrain. Fatigue life prediction from fe-safe using Abaqus results (d)in the section and (e) on the surface with minimum predicted life of15,847 cycles indicated.


with the smallest predicted number of cycles to failure islocated on the surface of the specimen. In the following,the fatigue life prediction for a specimen is always takenas the shortest life resulting at any node in the fe-safecalculations.The measured and predicted fatigue lives for all 25

porous specimens are compared in Figure 7. A line ofperfect correspondence is provided in the gure, to helpdetermine whether a prediction is conservative (belowthe line) or nonconservative (above the line). It can beseen that 18 of the predicted fatigue lives are within afactor of 10 of the test results, which can be regarded asgood agreement. Of these, only ve are conservative.Seven data points are in poor agreement with themeasurements (i.e., by more than a factor of 10) and arenonconservative, and two of these have predicted livesthat are approximately three orders of magnitude longerthan the measurements. While these results are generallyencouraging, the overall nonconservative nature of thepredictions warrants further investigation.It is well known that fatigue life predictions are very

sensitive to local stress concentrations. This issue is ofparticular importance in the present study, because thesimulations rely on pore fractions dened over avolume that is large compared to the microscopic poregeometry. It is unlikely that a node spacing of 0.25 mmresolves the local stress concentrations around smallshrinkage pores, which can result in nonconservativefatigue life predictions if the pores are located in ahighly stressed region. Therefore, a mesh sensitivitystudy was performed. Specimen 19 was simulated usingthree FEA meshes that are ner than the baseline grid(0.25 mm node spacing). The three ner meshes have

node spacings of 0.16, 0.12, and 0.09 mm. Figure 8shows the predicted stress elds on the surface ofspecimen 19, for all four meshes. The predictedmaximum stress increases from 400 MPa for thebaseline mesh to 467, 488, and 602 MPa for the threener meshes. Note that the applied stress amplitude forspecimen 19 is 126 MPa. Obviously, higher maximumstresses result in lower fatigue lives. The predictedfatigue lives for the four meshes are indicated assquares in Figure 7. The fatigue life predictions varyfrom approximately 160,000 cycles for the coarsestmesh to 8700 for the nest mesh. For specimen 19, theresult for the nest mesh is in good agreement with themeasured fatigue life of 13,013 cycles to failure. Asimilar mesh renement study was performed forspecimen 20, where the baseline mesh gives a fatiguelife prediction that is three orders of magnitude abovethe measured life. In addition to the four meshes usedfor specimen 19, simulations were also performed withtwo meshes that are coarser than the baseline mesh(node spacings of 0.42 and 0.58 mm). The predictedfatigue lives for specimen 20, shown in Figure 7 assquares, strongly decrease with increasing mesh ne-ness. The coarsest mesh yields a fatigue life predictionof approximately 3.4 9 109 cycles to failure. The threenest meshes give approximately the same fatigue life,indicating that the predictions converge to a constantvalue for a suciently ne mesh. However, theconverged value of approximately 4 9 106 cycles tofailure does not agree well with the measured life of250,000 cycles for specimen 20. This disagreement maybe related to the limited sensitivity of the lm radio-graphs from which the porosity eld was reconstructed,as is investigated in more detail in Section IV. Becausethe resolution of the radiography was approximately100 lm, a node spacing that is smaller than thisvalue does not change or improve the fatigue lifepredictions.Mesh renement studies conducted for other speci-

mens (such as shown in Figure 7 for specimen 17)showed the same trend. Generally, for those specimensfor which the prediction for a coarse mesh is alreadyclose to the measurement, further mesh renement doesnot decrease the predicted fatigue life substantially. Forthose specimens for which the baseline prediction isvastly nonconservative (e.g., specimen 20), the meshrenement has a much stronger eect. These dierencescan be attributed to the nature of the porosity distribu-tion in the specimens. Computational meshes with nodespacing on the order of 100 lm are generally notpractical in the structural analysis of entire metalcastings. Even for the relatively small specimens of thepresent study, such a ne mesh pushes the limits ofcurrent computational capabilities. Furthermore, itappears from the mesh renement study for specimen20 that porosity features smaller than 100 lm can playan important role in the fatigue behavior of thespecimens. Therefore, in the following, an approximatesimulation method is devised that reduces the strongmesh dependence of the results shown in Figure 7 andpredicts the measured fatigue lives more closely, even fora relatively coarse mesh.

Arrows indicates direction ofincreasing mesh fineness

Specimen 20

Specimen 19

Specimen 17

Measured Fatigue Life, Nf (Cycles to Failure) 102

102

103

104

105

106

107

108

109

1010

103 104 105 106 107

Pred

icte

d Fa

tigue

Life

, Nf (C

ycles

to Fa

ilure)

Fig. 7Comparison between measured and predicted fatigue lives ofspecimens with additional runs made for specimen 19 at three nergrids, and at two coarser and three ner nite element mesh gridsfor specimens 17 and 20.


IV. ADAPTIVE SUBGRID MODELFOR FATIGUE LIFE PREDICTION

The adaptive subgrid model developed in the presentstudy is based on the idea of using a local fatigue notchfactor at each node in the fatigue life calculations. Thisnotch factor is intended to account for the eect ofporosity that is not at all or only partially resolved in the

simulations. The lack of resolution can stem from eithera nite element mesh that is too coarse or a porosity thatis too small to be detected by standard lm radiography(such porosity is referred to as microporosity in thediscussion that follows). The use of a local fatigue notchfactor is dierent from the common practice of applyingthe same notch factor to an entire part or the entire

Fig. 8Predicted axial stress distribution on surface for four grids used to predict the fatigue life of specimen 19 shown in Fig. 7. Finer gridsgive higher stresses from more detailed porosity eld.


surface of a part.[3,13] In the present model, a fatiguenotch factor is only applied at those locations of the partat which the stress concentrations due to porosity arenot fully resolved. The magnitude of the fatigue notchfactor is calculated as a function of the local nodespacing and the local pore size. If the node spacing issmall enough that the stress concentration around apore of a certain size is fully resolved, the local fatiguenotch factor would be equal to unity. If the pore is sosmall relative to the node spacing that the FEA does notpredict any stress concentration, the local fatigue notchfactor would be equal to some maximum value thatcorresponds to that pore. This aspect is what makes thepresent subgrid model adaptive. The development of theadaptive subgrid model involves numerous consider-ations that are discussed separately in Sections Athrough E.

A. Fatigue Notch Factor

A fatigue notch factor, Kf, is commonly used infatigue life calculations to account for the eects ofnotches or discontinuities.[3] The fatigue notch factorrelates the local notch root stress and strain amplitudesto the nominal true stress and strain amplitudes througha relationship known as Neubers rule; i.e.,

De2

Dr2 K2f

De2

DS2

6

where Dr/2 and De/2 are the local axial stress andstrain amplitudes at the notch root, respectively, andDS/2 and De/2 are the nominal true stress and strainamplitudes, respectively. The multiaxial version ofNeubers rule uses the stresses and strains in the criti-cal plane, as explained previously in connection withthe BrownMiller algorithm. Equation [6] is solvedtogether with the cyclic stress-strain curve for thenotch stress and strain amplitudes; i.e.,

De2 Dr2E

Dr2K0

1n0

7

The notch strain amplitude resulting from the simulta-neous solution of Eqs. [6] and [7] is then used in thestrain-life equation (i.e., Eq. [5]) to calculate the fati-gue life. The fatigue notch factor, Kf, in Eq. [6] isobtained from[3]

Kf 1 Kt 11 a=r 8

where Kt is the stress concentration factor and r is thenotch root radius in millimeters. The material constanta is obtained from the following relation originallydeveloped for wrought steel:[44]

a 0:0254 2070Su

1:89

where Su is the ultimate strength in MPa (Table II). Forr a, which for Su = 1144 MPa is approximately thecase for r> 1 mm, Eq. [8] yields that Kf = Kt and thematerial is said to be fully notch sensitive.[3]

The use of the fatigue notch factor concept outlinedhere in a strain-life calculation results in accuratepredictions of the eect of microporosity on fatigue ofcast steel.[19] In Reference 19, additional fatigue tests arepresented with cast steel specimens that have onlymicroporosity. Microporosity is generally thought of asporosity that is not detected in standard lm radiogra-phy. In the context of the present measurements, it isdened as porosity with a pore volume fraction less thanapproximately 1 pct and with pore radii no larger thanapproximately 100 lm. Such microporosity should beopposed to the macroporosity in the present specimens(Table I). In the microporosity fatigue life calculationsof Reference 19, the notch root radius, r, was takenequal to the radius of the pore, Rp, found on the fracturesurface to be responsible for failure, i.e., r = Rp. Thestress concentration factor, Kt, was taken as a rstapproximation to be equal to 2.045, which is the valuecorresponding to a single spherical hole in an essentiallyinnite body.[45]

Because the fatigue notch factor model discussed herewas found to work well for microporosity, it forms thebasis for the present subgrid model. As is shown inSection IVB, pores with a radius less than 100 lm arenot at all resolved for nite element node spacingsgreater than approximately 0.2 mm (in fact, nodespacings less than 10 lm would be necessary to correctlycalculate the stress concentration for a 100-lm radiuspore). For such pores and meshes, the full fatigue notchfactor as given by Eq. [8] should thus be applied in thefatigue life calculations. For fully unresolved pores, as inthe limiting case of microporosity, the simulations canthen be expected to yield the correct fatigue life.

B. Adaptive Stress Concentration Factor

As discussed in Section A, when performing a niteelement stress analysis, and the mesh is so coarse relativeto the pore size that the predicted stresses are notaected by the presence of the pore (as would beexpected for microporosity), the stresses at the locationof the pore should be enhanced by a factor that gives thecorrect stress concentration corresponding to the pore.This factor is referred to as a stress concentration factor,Kt, and values for holes of various shapes and sizes canbe found in handbooks (e.g., Reference 45). However, ifthe node spacing is of the same order as or smaller thanthe pore size, a nite element stress analysis wouldpredict some or all of the enhanced stresses due to thepresence of the pore. In that case, the factor used toenhance the stresses should be less than the full stressconcentration factor from a handbook or even equal tounity. Thus, the objective of this section is to determinea relation for an adaptive stress concentration factor,Kt,a, that is a function of the node spacing, Ln, relative tothe pore size. For simplicity, pores are approximated assingle spherical holes of a certain eective radius, Rp, inan essentially innite body, so that the maximum valuefor Kt,a is equal to 2.045.

[45]

A nite element elastic stress analysis was performedof a single spherical hole with Rp = 0.1 mm inside of a5-mm-diameter axially loaded cylinder made of 8630


steel. This cylinder diameter is large enough that thepredicted stresses around the hole are not inuenced bythe boundaries. The hole is modeled as a spherical eldof nodes having 100 pct porosity (/ = 1); the hole isnot modeled as a feature in the nite element mesh.The elastic properties at each node are calculated fromEqs. [3] and [4], such that the elastic modulus, E, isessentially equal to zero at the nodes corresponding tothe hole. The applied load on the cylinder ends is takento be 207.4 MPa. Because Kt = 2.045 for a singlespherical hole,[45] the maximum stress at the hole shouldbe equal to 424 MPa. Figure 9 shows the predictedstress elds at a midsection cut of the hole for threedierent node spacings: 0.025, 0.1, and 0.2 mm. For thesmallest node spacing, the stresses around the hole arewell resolved and the predicted maximum stress is equalto 411 MPa. This stress is very close to the handbook

value of 424 MPa, so that Kt,a 1. For the interme-diate node spacing, the hole is poorly resolved and themaximum stress from the FEA is equal to only264 MPa. Hence, the corresponding adaptive stressconcentration factor is Kt,a = 424/264 = 1.606. Forthe coarsest mesh, only a single node is present withinthe hole and the predicted maximum stress from theFEA of 219 MPa is close to the nominal applied stressof 207.4 MPa; therefore, Kt,a = 424/219 = 1.94.These results can be generalized by realizing that, for

a single hole in an innite body, the full stressconcentration factor is independent of the radius ofthe hole. Figure 10 shows the computed adaptive stressconcentration factor, Kt,a, as a function of the nodespacing to the pore radius ratio, Ln/Rp. It was veriedthat the same result is obtained for hole radii other than0.1 mm. It can be seen from Figure 10 that, forLn/Rp< 0.08, the adaptive stress concentration factoris equal to unity, implying that the stresses around thehole are fully resolved. For Ln/Rp> 4, the mesh is socoarse that the full stress concentration factor of 2.045must be applied. For use in the present simulations, thefollowing curve was t through the computed datapoints

Kt;a 2:045 1 2:045 exp0:08 Ln

Rp

1:286

10

for Ln/Rp> 0.08. Equation [10] then provides the stressconcentration factor for use in Eq. [8]. For discontinu-ities other than a single spherical hole, a value that isdierent from 2.045 could be used in Eq. [10]. Becausethe use of Kt = 2.045 yields accurate fatigue life predic-tions for the microporosity specimens in Reference 19,this value is kept here.

C. Implementation in fe-safe

The software fe-safe only allows for the application ofa constant fatigue notch factor to the entire surface of

Fig. 9Predicted maximum principal stress distribution for a100 lm radius spherical hole in a 5 mm diameter cylinder under aloading of 207 MPa, with increasing mesh neness from (a) through(c). Node spacing and maximum stress for each case provided.

1

1.2

1.4

1.6

1.8

2

2.2

0 1 2 3 4 5 6Node Spacing/Hole Radius, L n /R p

100 m radius hole

Ada

ptiv

e St

ress

Con

cent

ratio

n Fa

ctor

, Kt,a

Curve fit: Equation (11)

Fig. 10Adaptive stress concentration factor Kt,a required toachieve correct fatigue life at ve node spacing to hole radius ratiosfor a 100 lm radius spherical hole modeled using a porosity eld.


the part or the entire part. However, the use of thepresent adaptive subgrid model necessitates the use of afatigue notch factor that varies from node to node. Thisissue can be resolved by making use of the fact thatfe-safe allows for temperature dependent fatigue prop-erties (r0f; e

0f; b, and c) in the strain-life equation, Eq. [5].

That feature is utilized in the present implementation ofthe adaptive subgrid model, by importing the calculatedfatigue notch factors at each node as a temperature eldinto fe-safe, and evaluating the fatigue properties as afunction of the fatigue notch factor. The functionaldependence between the fatigue properties and thefatigue notch factor must be such that the strain-lifeequation gives the same life with the modied(Kf dependent) fatigue properties as with the true fatigue

properties but the strains being the notch strains fromNeubers rule, Eq. [6].To determine the fatigue properties as functions of

Kf, software was developed that minimizes the errorbetween a strain-life curve generated for a given Kfvalue and the true fatigue properties from Table III,and a strain-life curve with Kf = 1 but modied fatigueproperties. For a given Kf, the fatigue properties r0f, e

0f,

b and c are iterated upon and solved for until the twostrain-life curves match. The resulting dependencies ofe0f, b, and c on Kf are plotted in Figures 11(a), (b), and(c), respectively. The present minimization proce-dure revealed that r0f should not be made a functionof Kf (and thus be kept at the true value given inTable III).

D. Pore Size Model

The sole remaining unknown in the adaptive subgridmodel is the pore radius, Rp. It is needed in Eq. [8] forthe fatigue notch factor and in Eq. [10] for the adaptivestress concentration factor. Porosity is always charac-terized by a distribution of pore sizes, rather than asingle pore radius, in the volume element over whichthe porosity volume fraction is dened (e.g., References27 through 29). For the purpose of fatigue lifecalculations it is important to know the radius of thelargest pore in the distribution, because it is that porethat initiates failure.[19,24,2729,34] Thus, the present poresize model is concerned primarily with predicting themaximum pore radius, Rp,max. In order to keep themodel relatively simple, the maximum pore radius isassumed to be dependent on the pore volume fraction,/, only. As explained here, results from both aprobabilistic model of pore growth and merging andfrom experimental measurements are used to developthe pore size model.The following example illustrates the importance of

distinguishing between the mean and maximum poreradii. In Reference 19, the present authors andco-workers performed image analysis on polished metal-lographic sections of 8630 cast steel specimens havingmicroporosity. The average pore volume fraction wasfound to be / = 0.7 pct, and the mean distancebetween the pores, Lp, was observed to range from 177to 344 lm. The pore number density, n (i.e., the numberof pores per unit volume) can then be estimated from therelation n L3p to be between 2.4 9 1010 and1.8 9 1011 pores/m3. Assuming the pores are sphericaland have a uniform radius, the pore volume fraction isgiven by

/ n 4p3R3p 11

Using these values for / and n, a mean pore radius ofRp = 21 to 41 lm is obtained from Eq. [11], which wasveried by the image analysis. However, images of thefracture surface of the microporosity specimens revealedthat failure always initiated from much larger, isolatedpores with a radius on the order of 100 lm.[19] Thisvalue corresponds to the radius of the largest pore.

-0.22

-0.20

-0.18

-0.16

-0.14

-0.12

Fatigue Notch Factor

0.050.100.150.200.250.300.350.400.45

Fatigue Notch Factor

-0.78

-0.76

-0.74

-0.72

-0.70

-0.68

1.0 1.5 2.0 2.5

1.0 1.5 2.0 2.5

1.0 1.5 2.0 2.5Fatigue Notch Factor

(a)

(b)

(c)

Fatig

ue D

uctil

ity C

oeffi

cien

t f

Fatig

ue D

uctil

ity E

xpon

ent c

Fa

tigue

Stre

ngth

Exp

onen

t b

Fig. 11Fatigue properties as functions of notch factor for (a) fati-gue strength exponent b, (b) fatigue ductility coecient ef, and (c)fatigue ductility exponent c.


Taking Rp = 100 lm and n as estimated earlier, Eq. [11]yields pore volume fractions, /, ranging from 10 to75 pct, which is clearly unrealistic. Thus, Eq. [11] is notvalid for the maximum pore radius.As part of the present study, a probabilistic pore size

distribution model was developed to better understandthe phenomena of pore growth and merging and therelationship between the maximum and mean pore radii.Due to space limitations, it is only briey described here.The pores are assumed to nucleate instantaneously witha certain initial volume fraction and number density anda lognormal size distribution. Based on the micropo-rosity measurements mentioned earlier, the shape ofthe initial size distribution is adjusted such that for/ = 0.7 pct and n = 9.97 9 1010 m3, the mean poreradius, Rp, is equal to 26 lm and the maximum poreradius, Rp,max, in the distribution is 100 lm. Eachpore in the distribution is assumed to grow at a rateproportional to the square of its radius. In addition, apore merging model is implemented. Pores are assumedto be capable of merging if their radius exceeds half ofthe mean spacing between the pores. The frequency ofpore merging is determined by the probability that twopores will take part in a merging event. This probabilityis dependent on the product of the number density ratiosof the merging pores (either within a bin of a single poresize meeting the merging criterion or between two binsof dierent pore sizes). The number density ratio is thenumber density of one pore size bin divided by the totalnumber density of pores. Merged pores continue to formand grow along with the distribution. Due to merging,the total number density of pores continually decreasesas the mean pore radius and total porosity volumefraction increase.Representative calculated pore size distributions are

shown in Figure 12. In Figure 12(a), the curve corre-sponding to / = 0.1 pct represents the initial pore sizedistribution. With an increasing pore fraction, themaximum of the distribution shifts to higher pore radiiand the distribution becomes narrower. At approxi-mately / = 3 pct, signicant pore merging is starting tooccur. This can be seen in Figure 12(a) by the area underthe distribution decreasing and, hence, the total porenumber density decreasing. In addition, the pore sizedistributions develop a tail at pore radii that are muchlarger than the mean. This can be better seen inFigure 12(b), where the predicted pore size distributionfor / = 10 pct is shown in a log-log plot. At 10 pct, thetotal number density, mean pore radius, and maximumpore radius are approximately given by n = 1.17 9106 m3, Rp = 0.73 mm, and Rp,max = 2.61 mm,respectively. The maximum pore radius is dicult todetermine from the distribution because of its long tail,but for / = 10 pct, the value of 2.61 mm correspondsto the mean of the merged pores (Figure 12(b)). Thismaximum pore radius agrees with the measurements ofASTM radiographs for shrinkage porosity in steelcastings presented in Reference 46.For use in the present adaptive subgrid model, the

maximum pore radius results from the probabilistic poresize distribution model were t to the following piece-wise function:

Rp;max 2825:2/1:0 25; 846:63/ 275; 154:35/2for/ 0:0386

Rp;max 1:76888 1:66246 exp 21; 223:91/3:98865

for 0:0386

provide a rst-order correction to the strong meshdependence of the fatigue life simulations, the maximumpore radius model can be considered suciently accu-rate. In summary, the function for Rp,max given here isused for both the notch root radius, r, in Eq. [8] for thefatigue notch factor, and the pore radius, Rp, in Eq. [10]for the adaptive stress concentration factor.

E. Results Using the Adaptive Subgrid Model

The adaptive subgrid fatigue life model is rst testedfor specimen 20. As previously shown in Figure 7, theoriginal fatigue life predictions for specimen 20, withoutthe subgrid model, are highly dependent on the nodespacing. A comparison of those results with the corre-sponding predictions in which the adaptive subgridmodel is employed in the simulations is shown inFigure 14. For the three nest meshes (0.16, 0.12, and0.09 mm node spacing), the fatigue life predictions arethe same with and without the subgrid model; further-more, the results are independent of the node spacingand can thus be considered converged. The convergedvalue for the predicted fatigue life is ~4 9 106 cycles tofailure. The agreement illustrates that the presentadaptive subgrid model correctly reduces to a standardstress/fatigue life simulation for a suciently ne niteelement mesh. In other words, the introduction of thefatigue notch factor in the adaptive subgrid model doesnot reduce the predicted fatigue life to a value below theone for a fully resolved FEA calculation. For the threecoarser meshes (0.25, 0.42, and 0.58 mm node spacing),the fatigue life predictions are substantially above theconverged result of approximately 4 9 106 cycles tofailure. However, the fatigue lives obtained with theadaptive subgrid model are consistently closer to theconverged value and are more mesh independent. Forthe coarsest mesh, the adaptive subgrid model has thestrongest benecial eect and reduces the fatigue lifepredicted without the subgrid model by almost twoorders of magnitude. While these results for specimen 20are encouraging and indicate that the adaptive subgridmodel is working as intended, it can be seen fromFigure 14 that, for the three coarser meshes, the fatigue

lives predicted with the subgrid model are still up to 1order of magnitude larger than the converged value.This nding suggests that a larger maximum stressconcentration factor in Eq. [10], instead of the 2.045value, would produce even better results. While that iscertainly true, the use of a larger maximum stressconcentration factor in the adaptive subgrid model isnot desirable, because the value of 2.045 gave accuratefatigue life predictions for the microporosity specimensin Reference 19. Furthermore, note that even the nestmesh results are still considerably above the measuredfatigue life for specimen 20 (~250,000 cycles to failure(Table I)); thus, factors other than the mesh resolutionare causing disagreement between the measured andpredicted fatigue lives.Figure 15 shows the eect of the adaptive subgrid

model on the fatigue life predictions for all 25 spec-imens. These simulations were performed with a nodespacing of 0.25 mm. The vertical lines are used toindicate in the gure the movement of the predictionsdue to the use of the subgrid model. Relative to themeasurements, the predictions are much improved,especially for the highly nonconservative data points.It is interesting that those predictions that were alreadyclose to the measurements are not aected by thesubgrid model and the good agreement remains. This isespecially true for the lower lived specimens with fatiguelives of less than ~10,000 cycles to failure. For thosespecimens, failure is controlled by large porosity fea-tures and the stress eld is already adequately resolvedwith a node spacing of 0.25 mm.Although the adaptive subgrid model results in a

considerable improvement in the fatigue life predictions,there are still six data points in Figure 15 for which theprediction is more than an order of magnitude above themeasurement. This could be the result of the porosityeld not being suciently resolved in the X-ray tomo-graphy. In particular, the radiography does not detectmicroporosity, because the resolution is limited by theradiographic lm sensitivity to 100 lm, as mentioned

0.0

0.5

1.0

1.5

2.0

2.5

3.0

0.00 0.05 0.10 0.15 0.20 0.25 0.30

Max

imum

Por

e Ra

dius

, Rp,

max

(m

m)

Pore Volume Fraction, (-)

Start of significant pore merging

1 mm

1 mm

0.7% Porosity

7% Porosity

10% Porosity

20% Porosity

1 mm

Fig. 13Relationship for maximum pore size vs volume fractioncompared with observed porosity in specimens (5 mm diameter).

0.0 0.1 0.2 0.3 0.4 0.5 0.6Node Spacing (mm)

Without adaptive sub-grid model

With adaptive sub-grid model

1010

109

108

107

106

105Pred

icte

d Fa

tigue

Life

,Nf(C

ycles

to Fa

ilure)

Fig. 14Eect of nite element model node spacing on predictedfatigue life from fe-safe for specimen 20. Original method withoutsubgrid fatigue model and with subgrid model with Kf,max = 2.045.


earlier. Microporosity was always observed on metallo-graphic sections, such as those shown in Figure 2, inregions next to larger pores. This can be important in afatigue life calculation, because high stress concentra-tions occur next to large pores. In order to investigatethe sensitivity of the predictions to the presence ofmicroporosity, simulations were performed with auniform background microporosity having a maxi-mum pore radius, Rp,max, of 100 lm. This is easilyaccomplished in the present subgrid model by modifyingthe pore size model, Eq. [12], such that all values ofRp,max that are less than 100 lm are overwritten with100 lm (which only occurs for / less than 0.1 pct). Theporosity eld used in the nite element stress analysis ofthe specimens was not changed, because microporositycorresponds to very small pore fractions with a negli-gible eect on the elastic properties.The eect of the background microporosity on the

fatigue life predictions for all 25 specimens is shown inFigure 16. The simulation results in this gure wereobtained using the adaptive subgrid model and a nodespacing of 0.25 mm. Here, the vertical lines indicate themovement of the predictions due to the addition of thebackground microporosity. As expected, the addition ofthe background microporosity lowers the fatigue lifepredictions for all specimens. However, the magnitudeof the eect is not the same for all specimens. For somespecimens, background microporosity has a negligibleeect, while for others, the predicted fatigue life changesby almost two orders of magnitude. These dierencescan be attributed to the nature of the porosity eld in

the specimens and the resulting stress redistributions.Overall, with the addition of the background micropo-rosity in the simulations, the agreement between themeasured and predicted fatigue lives can now beconsidered satisfactory.

V. CONCLUSIONS

A simulation method is developed for predicting theeect of shrinkage porosity on the fatigue life of steelcastings. The method is validated using previouslymeasured fatigue life data for 25 cast steel test speci-mens containing up to 21 pct porosity in the gagesection.[19] The simulation method relies on the knowl-edge of the three-dimensional porosity eld. In thepresent study, the porosity eld is obtained from X-raytomography with a resolution of approximately100 lm.[1] After importing the porosity eld into theFEA software, an elastic stress analysis of the fatiguetests is conducted via the method developed in Reference1. Using this method, the local elastic properties arereduced according to the volume fraction of porositypresent at an FEA node. The computed stresses areimported into fatigue analysis software to calculate thefatigue life distribution in the specimens using the strain-life equation for sound steel. It is found that themeasured fatigue lives of several of the specimens testedare vastly overpredicted when a nite element mesh isused that is too coarse to resolve all of the porosity-induced local stress concentrations. For this reason, anadaptive subgrid model is developed that employs alocal fatigue notch factor to approximately account for

Without adaptive sub-grid model With adaptive sub-grid model

Measured Fatigue Life, Nf (Cycles to Failure)

Pred

icte

d Fa

tigue

Life

, Nf (C

ycles

to Fa

ilure)

102 103

103

102

104

105

106

107

108

109

104 105 106 107

Fig. 15Comparison between measured and predicted fatigue livesof specimens for node spacing 0.25 mm for original method(no Kf) and for the subgrid fatigue model with Kf,max = 2.045. Pre-diction uses Abaqus simulated stress eld and subgrid fatigue modelwith fatigue properties dependent on Kf, as shown in Fig. 11. Multi-axial BrownMiller algorithm with Morrow mean stress correction isused in fe-safe.

Pred

icte

d Fa

tigue

Life

, Nf (C

ycles

to Fa

ilure)

With adaptive sub-grid model, no background microporosity

With adaptive sub-grid model, with backgroundmicroporosity, Rp = 100 m

Measured Fatigue Life, Nf (Cycles to Failure) 102

102

103

104

105

106

107

108

109

103 104 105 106 107

Fig. 16Comparison between measured and predicted fatigue livesof specimens for node spacing 0.25 mm for the subgrid fatiguemodel with Kf,max = 2.045 (higher values), and the same model butusing a uniform background pore size of 100 lm radius (even atnodes at which tomography gives material as 100 pct sound).


the eect of under-resolved porosity. In this subgridmodel, an adaptive stress concentration factor is calcu-lated that depends on the size of a shrinkage porerelative to the node spacing of the nite element mesh.The maximum pore radius is obtained from a probabi-listic pore size distribution model as a function of thelocal pore volume fraction. With the adaptive subgridmodel and uniform background microporosity with amaximum pore radius of 100 lm, the measured andpredicted fatigue lives are found to be in good overallagreement, even for a relatively coarse FEA mesh. Thepresent simulation method is suciently general andexible that it can be used for the fatigue analysis ofcomplex-shaped steel castings containing porosity. Forsuch production steel castings, X-ray tomographicreconstructions of the three-dimensional porosity eldare generally not available. It is anticipated that theporosity eld can instead be predicted using advancedcasting simulation software.[47,48] A preliminary casestudy demonstrating such an integrated simulationmethodology was presented in Reference 2.

ACKNOWLEDGMENTS

This research was undertaken through the AmericanMetalcasting Consortium (AMC), which is sponsoredby the Defense Supply Center Philadelphia (DSC)(Philadelphia, PA) and the Defense Logistics Agency(DLA) (Ft. Belvoir, VA). This work was conductedunder the auspices of the Steel Founders Society ofAmerica (SFSA) through substantial in-kind supportand guidance from SFSA member foundries. Anyopinions, ndings, conclusions, or recommendationsexpressed herein are those of the authors and do notnecessarily reect the views of DSC, DLA, or theSFSA or any of its members.

REFERENCES1. R.A. Hardin and C. Beckermann: Metall. Mater. Trans. A, 2007,

vol. 38A, pp. 29923006.2. R.A. Hardin, R.K. Hu, and C. Beckermann: in Modeling of

Casting, Welding and Advanced Solidication Processes XI, C.Gandin and M. Bellet, eds., TMS, Warrendale, PA, 2006, pp. 65360.

3. R.I. Stephens, A. Fatemi, R.R. Stephens, and H.O. Fuchs: MetalFatigue in Engineering, 2nd ed., Wiley-Interscience, New York,NY, 2000.

4. 2002 Annual Book of ASTM Standards, ASTM, West Cons-hohocken, PA, 2002, vol. 03.01, pp. 56983.

5. S.S. Manson: Behavior of Materials under Conditions of ThermalStress, NACA TN2933, National Advisory Committee forAeronautics, Washington, DC, 1954.

6. L.F. Con: Trans. ASME, 1954, vol. 76, pp. 93150.7. D.F. Socie and G.B. Marquis: Multiaxial Fatigue, Society of

Automotive Engineers, Warrendale, PA, 2000.8. B.-R. You and S.-B. Lee: Int. J. Fatigue, 1996, vol. 18 (4), pp. 235

44.9. C.-C. Chu: Int. J. Fatigue, 1996, vol. 19 (Suppl. 1), pp. S325S330.10. D. Taylor, P. Bologna, and K. Bel Knani: Int. J. Fatigue, 2000,

vol. 22, pp. 73542.11. Y. Liu, B. Stratman, and S. Mahadevan: Int. J. Fatigue, 2006,

vol. 28, pp. 74756.12. Y. Liu, B. Stratman, and S. Mahadevan: Rel. Eng. Sys. Saf., 2008,

vol. 93, pp. 45667.

13. fe-safe User Manual, Safe Technology, Ltd., Sheeld, UK, 2006,pp. 23339.

14. P.C. Paris and F. Erdogan: Trans. ASME, J. Basic Eng., 1963,vol. D85, pp. 52834.

15. S.C. Haldimann-Sturm and A. Nussbaumer: Int. J. Fatigue, 2008,vol. 30, pp. 52837.

16. P. Baicchi, L. Collini, and E. Riva: Eng. Fract. Mech., 2007,vol. 74, pp. 53948.

17. Y. Nadot and V. Denier:Eng. Fail. Anal., 2004, vol. 11, pp. 48599.18. P. Hausild, C. Berdin, P. Bompard, and N. Verdiere: Proc. 6th

World 2000 Duplex Conf., Associazione Italiana di Metallurgia,Venice, Italy, 2000, pp. 20918.

19. K.M. Sigl, R. Hardin, R.I. Stephens, and C. Beckermann: Int. J.Cast Met. Res., 2004, vol. 17 (3), pp. 13046.

20. S. Jayet-Gendrot, P. Gilles, and C. Migne: Fatigue and Fracture:1997 PVP-Vol. 350, 1997, ASME, vol. 1, pp. 10716.

21. M. Kohno and M. Makioka: AFS Trans., 1970, vol. 78, pp. 916.22. K. Chijiwa, T. Nakayama, and M. Imamura: 35e CIF, vol. 36,

pp. 112.23. P. Heuler, C. Berger, and J. Motz: Fatigue Fract. Eng. Mater.

Struct., 1992, vol. 16, pp. 11536.24. T. Billaideau, Y. Nadot, and G. Bezine: Acta Mater., 2004, vol. 52,

pp. 391120.25. Y. Nadot, J. Mendez, and N. Ranganathan: Int. J. Fatigue, 2004,

vol. 26, pp. 31119.26. T. Mansson and F. Nilsson: Int. J. Cast Met. Res., 2001, vol. 13

(6), pp. 37378.27. X. Zhu, J.Z. Yi, J.W. Jones, and J.E. Allison: Metall. Mater.

Trans. A, 2007, vol. 38A, pp. 111122.28. X. Zhu, J.Z. Yi, J.W. Jones, and J.E. Allison: Metall. Mater.

Trans. A, 2007, vol. 38A, pp. 112335.29. J. Linder, M. Axelsson, and H. Nilsson: Int. J. Fatigue, 2006,

vol. 28, pp. 175258.30. D.L. McDowell, K. Gall, M.F. Horstemeyer, and J. Fan: Eng.

Fract. Mech., 2003, vol. 70 (1), pp. 4980.31. C. Sonsino and J. Ziese: Int. J. Fatigue, 1993, vol. 15, pp. 7584.32. A. Dabayeh and T.H. Topper: Fatigue Fract. Eng. Mater. Struct.,

2000, vol. 23, pp. 9931006.33. J.-Y. Bue`re, S. Savelli, P.H. Jouneau, E. Maire, and R. Fouge`res:

Mater. Sci. Eng., A, 2001, vol. 316, pp. 11526.34. Q.G. Wang and P.E. Jones: Metall. Mater. Trans. B, 2007,

vol. 38B, pp. 61521.35. Q.G. Wang, D. Apelian, and D.A. Lados: J. Light Met., 2001,

vol. 1, pp. 7384.36. B. Skallerud, T. Iveland, and G. Harkegard: Eng. Fract. Mech.,

1993, vol. 44 (6), pp. 85774.37. Y. Murakami: Metal Fatigue: Eects of Small Defects and

Nonmetallic Inclusions, Elsevier Science, Ltd., Oxford, UnitedKingdom, 2002.

38. Y. Xue, C.L. Burton, M.F. Horstmeyer, D.L. McDowell, and J.T.Berry: Metall. Mater. Trans. B, 2007, vol. 38B, pp. 60106.

39. A.C. Kak and M. Slaney: Principles of Computerized TomographicImaging, Society for Industrial and Applied Mathematics,Philadelphia, PA, 2001, pp. 27596.

40. R.I. Stephens: Fatigue and Fracture Toughness of Five Carbon orLow Alloy Cast Steels at Room or Low Climatic Temperatures,Steel Founders Society of America, Des Plaines, IL, 1982.

41. IMSL Fortran Subroutines for Mathematical Applications: Math/LibraryVolumes 1 and 2, VisualNumerics, Inc.,Houston, TX, 1997.

42. A.P. Roberts and E.J. Garboczi: J. Am. Ceram. Soc., 2000, vol. 83(12), pp. 304148.

43. Abaqus/Standard, Abaqus, Inc., Providence, RI, 2006.44. Fatigue Design Handbook, Society of Automotive Engineers,

Warrendale, PA, 1968, vol. 4, p. 29.45. R.E. Peterson: Stress Concentration Factors, Wiley-Interscience,

New York, NY, 1974, p. 137.46. K.D. Carlson, S. Ou, R.A. Hardin, and C. Beckermann: Int. J.

Cast Met. Res., 2001, vol. 14, pp. 16983.47. K.D. Carlson, Z. Lin, R.A. Hardin, C. Beckermann, G.

Mazurkevich, and M.C. Schneider: in Modeling of Casting,Welding and Advanced Solidication Processes X, D.M. Stefanescu,J.A. Warren, M.R. Jolly, and M.J.M. Krane, eds., TMS,Warrendale, PA, 2003, pp. 295302.

48. K.D. Carlson, Z. Lin, and C. Beckermann: Metall. Mater. Trans.B, 2007, vol. 38B, pp. 541555.


Outline placeholderAbs1IntroductionExperimentsExperimentsExperimentsExperiments

Fatigue life predictionFatigue life predictionFatigue life prediction

Adaptive subgrid model for fatigue life predictionAdaptive subgrid model for fatigue life predictionAdaptive subgrid model for fatigue life predictionAdaptive subgrid model for fatigue life predictionAdaptive subgrid model for fatigue life predictionAdaptive subgrid model for fatigue life prediction

ConclusionsConclusionsConclusions

/ColorImageDict > /JPEG2000ColorACSImageDict > /JPEG2000ColorImageDict > /AntiAliasGrayImages false /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 150 /GrayImageDepth -1 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict > /GrayImageDict > /JPEG2000GrayACSImageDict > /JPEG2000GrayImageDict > /AntiAliasMonoImages false /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict > /AllowPSXObjects false /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputCondition () /PDFXRegistryName (http://www.color.org?) /PDFXTrapped /False

/Description >>> setdistillerparams> setpagedevice