You are on page 1of 9
Copyright ©) 1996 Elsevier Science Lid Printed in Great Britain, All righs reserved PI: S0013-7944(96)00076-8 (0015-7948/97 $17.00 + 000 PREDICTION OF FRACTURE TOUGHNESS USING ARTIFICIAL NEURAL NETWORKS (ANNs) A. SEIBI Department of Mechanical Engineering, Sultan Qaboos University, Muscat, Sultanate of Oman S. M. AL-ALAWI Department of Electrical & Electronics Engineering, Sultan Qaboos University, Muscat, Sultanate of ‘Oman Abstract—This paper explores the potential use of artificial neural networks (ANNs) in the field of frac- ture mechanics. It addresses the use of ANNs in predicting the average fracture toughness, K., of 7075- 651 aluminum alloy under uni- as well as biaxial loading at the room and higher temperatures. The fracture toughness prediction is based on the evaluation of critical values of J from experimental data Parameters that influence the value of the fracture toughness are used to develop the ANN model, and. the contribution of these variables to the variation of fracture toughness is then found, This paper also explores the effect of crack geometry, temperature and biaxiality on fracture toughness. Results indicate that ANN predicted the fracture toughness under different conditions with high accuracy. It also demonstrates that AINN is an excellent analytical tool that, if properly used, can reduce cost, time and ‘enhance structure reliability. Copyright © 1996 Elsevier Science Lid INTRODUCTION AMONG THE fracture failures responsible for interest in the field of fracture mechanics are those of pressure vessels, welded ships, gas transmission lines, large oil storage tanks and reactor vessels. In most cases, the origin of failure has been found at flaws which became critical as the stress was raised to a critical size while the member was subjected to operational stress. Although there have been many attempts to establish a theory which can explain the fracture process, most of them have failed to describe the fracturing mechanism accurately. Among the well-known and mostly used fracture criteria is the stress intensity factor approach developed first by Griffith{1] and then modified by Irwin 2] to establish a relationship between the critical strain energy release rate and the fracture toughness using the fracture mechanics theory. The recent developments of fracture mechanics as an analytical design tool made it possible for the designer to promote safe design structures by estimating the notch toughness of materials at service conditions. A large number of correlations of the fracture toughness, taken as a design parameter, for various crack geometries under different loading conditions are available in the literature[3-5]. Accurate estimation or prediction of the fracture toughness of a given material is essential in structural design to prevent catastrophic failure that is economically costly and may lead to the loss of human lives. To reduce cost and enhance reliability of a structure, a novel method for predicting the fracture toughness and examining the contribution of the variables that significantly affect the values of the fracture toughness is required. In this paper an artificial neural network (ANN) is proposed as the framework for develop- ing a model that predicts the fracture toughness of 7075-T651 aluminum alloy based on the im- portant parameters that influence the fracture toughness values. Artificial neural networks are computational models that are designed to mimic human brain architecture and operations [6]. These networks have shown remarkable performance when used to model complex linear and nonlinear relationships(7-9] and have been successfully applied to a variety of engineering pro- blems{10-12}, The ability of ANN to learn complex relationships between the input and output parameters, and to allow the user to examine the effect of each parameter on the fracture tough- ness once a model is developed, make it a significant analytical tool. Considerable savings in terms of cost and time could be obtained from using the trained ANN model. 31 312 ‘A, SEIBI and $. M. AL-ALAWI ‘Table 1. Dimensions of the three-point and anticlastic bending specimens Specimen ‘Width (in) Thickness (n.) Span (in) Beam 20 Lo. 80 Plate 30 01825 100 SYSTEM UNDER STUDY AND EXPERIMENTAL DESIGN Experimental data used to develop the ANN model were taken from ref.[13] where uni- as well as biaxial tests on beams and plates with straight and surface cracks were conducted. In this study, Baghat and Zamrik(13] used the J integral approach to predict the fracture toughness of plates under certain loading conditions by estimating the average value of the energy line integral, J, from experimental data using the concept of load-deflection curve for cracked and uncracked specimens. The average value of J is given by: 2 pin a f P dBcrack. a IL where P is the applied static load, Ay is the area of the remaining uncracked ligament, and dcrack is the deflection of the cracked specimens. This equation was used to estimate the average frac- ture toughness for cracked specimens given by Ei RK for plane stress 2 K, = VBI for plane stress ° Equations (2, 3) indicate that the fracture toughness depends on the crack geometry as well as the relative dimensions of the crack-to-the-specimen size, temperature and loading conditions. Equation (3) was used to estimate the fracture toughness of 7075-T651 aluminum alloy beam and plate specimens with embedded surface and straight cracks under uniaxial and biaxial loading. Different crack lengths and depths of a V shape with an angle of 60° were embedded on the specimens. The dimensions of both beam and plate specimens are included in Table 1. Table 2, however, contains the average dimensions of surface and straight cracks for the beam and the biaxial anticlastic bend specimens used during testing. The general procedure followed in ref. [13] during testing for data generation consists of first cycling the specimens between tensile and zero loads according to the ASTM E399-72|14] where the cycling frequency in all tests was maintained at 20 Hz. After a prescribed fatigue crack was generated, a tint-colouring dye was inserted into the crack to mark crack extension area when a specimen is subjected to static loading from which load-deflection curves were gen- erated. Once a crack extension due to the static load was obtained, the specimen was subjected to a second cycling procedure until fatigue failure took place and the specimen was broken into two separate pieces. In this procedure, a new fatigue crack is generated and the crack extension due to static load is bounded by two fatigue cracks. The crack extension area and crack depth were measured from the broken section of each specimen where A was evaluated using Table 2, Average dimensions of straight and surface eracks Specimen Surface crack length (in.) Surface crack depth (in.) Straight erack depth (in) Beam 075 0.260 0.260) 1.00 0310 0310 125 0.381 0381 1.50 0.450 0.450 175, 0.470 0.470 Plate 030 0.200 0.200 0.75 0.300 0.300 1.00 0.360 0.360 12s 0413 0.413 t 50 0.460 0.460 Artificial neural networks 313 Input Hidden Output layer layer layer Fig. 1. The architecture of the ANN model. Simpson's rule. Full details of the experimental set-up, procedures, specimen configurations and testing fixtures are included in ref.[13]. Design of the ANN model In this work, an ANN model is used to predict the fracture toughness of an aluminum alloy based on experimental data and developed relationships between K, and crack geometry, specimen dimensions and operating temperature. Figure 1 shows the multilayer feedforward per- ception artificial neural network used in this study. ANN architectural design As illustrated by Fig. 1, the network architecture is composed of many simple processing el- ements that are organized into a sequence of layers. These are the input layer, the hidden layer, and the output layer. The neurons in the input layer receive seven input signals representing the loading type used (LT) both uniaxial and biaxial, operating temperature (7), crack depth (a), crack length (2c), erack depth to crack length ratio (a/c), crack depth to thickness ratio (x), and plate length to crack length ratio (f); hence, seven neurons were used for input in the ANN architecture. The output layer, on the other hand, consists of two output neurons representing the average fracture toughness for surface and straight cracks. Between the input and output layers, generally, there is one or more hidden layers. Determining the number of hidden layers to use and the appropriate number of neurons to include in each hidden layer is not an exact science, Research in this area (15, 16] proved that one or two hidden layers with an adequate number of neurons are sufficient to model any sol- ution surface of practical interest. Furthermore, recent research findings show that the number of neurons to include in the hidden layer is a function of the number of training pairs available [17]. A large number of hidden layer nodes has a large number of associated undetermined par- ameters, and if the number of training pairs is small, the network will then tend to memorize rather than generalize. The author pointed out that in order to have a good approximation over the region of interest, an overdetermined network should be used. He suggested the following formula for calculating the appropriate number of hidden nodes to be used in a single hidden layer network if the number of training pairs is known: N = 6[Hn(In + 1) + On(Hn + 1)], @ where N = number of training pairs available 4a, constant greater than 1.0 (j.e., 0 = 1.25 would give a 25% overdetermined approxi- mation) ‘Hn = number of hidden nodes to be used in a network that has only one hidden layer In = number of input nodes ‘On = number of output nodes. 314 A. SEIBI and S. M. AL-ALAWI Using the above formula, Hn was found to be 2.98. This result indicates that three hidden nodes can be used in the hidden layer. This finding is validated by evaluating a number of hid- den nodes configurations and computing the root mean square error for both the design points and the test points. As a result of this evaluation, it was found that a network with three to five nodes on the hidden layer would be a good choice. The network containing four hidden nodes, however, yielded the smallest error over the region of interest. Based on these results, one hid- den layer containing four neurons is used to develop the ANN architecture. Data preparation Before the ANN model can be used to provide the desired output, the model has to be trained to recognize the relationships between the input parameters and the desired output. These relationships will be stored as connection weights between the different neurons. This pro- cess is called the training or the learning process. Prior to conducting the training process, a set of input-output patterns is first prepared. The set is developed through experimental work where the effect of changing input parameters on the fracture toughness of straight and surface cracks is examined, such as the type of load- ing, operating temperature, crack depth, etc. The input and output parameter ranges used in this study are shown in Tables 1, 2 and 3, Patterns within these ranges are evenly distributed so that the training can cover all possible typical load values. If this is not the case, training will tend to focus on regions where training patterns are densely clustered and neglect those that are sparsely populated, hence, producing inaccurate gains. Examples of the different training pat- terns used in the present work as part of the training data set are shown in Table 3 Network training ‘The multilayer feedforward network used in this work is trained using the Backpropagation (BP) paradigm developed in ref.[18]. The BP algorithm uses the supervised training technique. In this technique, the interlayer connection weights and the processing element thresholds are first initialized to small random values. The network is then presented with a set of training pat- terns, each consisting of an example of the problem to be solved (the input) and the desired sol- ution to this problem (the output), These training patterns are presented repeatedly to the ANN model, and weights are adjusted by small amounts that are dictated by the general delta rule [18]. This adjustment is performed after each iteration when the network’s computed output is different from the desired output. This process continues until weights converge to the desired error level or the output reaches an acceptable level. The following system of equations[19] pro- vides a generalized description of how the learning process is performed by the BP algorithm. ‘After initializing the connection weights between the input layer and the hidden layer, the connection weights between the hidden layer and the output layer, and the processing element (PE) thresholds, the following calculations are performed on each pattern pair (4y, C,) in the training set: Firstly, the input vector (4,) is presented to the PEs in the input layer where the activations of these PEs are filtered, and the activation values for the hidden layer are calculated using the equation bi =S (Yawn +6), (5) where a, is the output from the processing element / and f{) is the logistic function ‘Table 3. Examples of input and output training patterns used in developing the ANN mode! Parameter Input ‘Input ‘Input ——‘Input_—‘Input_—‘Input_—‘ Input-Output Output Case # IT TCC ain) ein) — ale a BR Surface K. Straight 4 1 25 047) 17505370470 —«d.140.—=«30:163——28.600 1 1 140 031 100 06200310 2.000 «47810 40.140 15 2 140 030 075 0.880 ©0364 = 4.000 46.720 35.780 20 1 20 0.38 125 0.609 ©0381 = 1.600 27.260 25.200 3 2 220 036 100 0.700436. 3.000 27.040 24.470 2 1 300 026 075 ©0693 0.260 © 2.670 17.240 20.280 35 2 300 046 150 0610 058721000 15.120 13.600 Artificial neural networks 315 f@sdtery! (6) Secondly, the hidden layer activations are filtered through the connections to the output layer using the equation o-s (Somes . 5) Oy ot and the error between the computed and the desired output layer processing element values is calculated using the equation 4 =o GF - 4). @) where d; is the output layer PE’s computed error and cis the output from the processing el- ement j. The error for each hidden layer PE relative to each d, can then be calculated using the equation ‘ €, = bill — bY wy dy o = where e; is the computed error for the hidden layer processing element i. Once all errors for the neurons in the hidden layer are calculated, the connections between the hidden layer and the ‘output layer can then be adjusted using the learning rate. Awy = absd). (0) ‘The output layer thresholds can also be adjusted using the equation AT; = adj. ay Once all connections between the output layer, the hidden layer, and the thresholds are adjusted, then the connections between the input layer and the hidden layer also need to be adjusted using the equation AV ig = Bares, (12) where Vj; is the amount of change made to each connection linking the PEs between the input layer and the hidden layer, and is a positive constant controlling the learning rate. Finally, the hidden layer thresholds are also adjusted using the equation AO; = Bey (13) This process is repeated until the error correction value d; is either sufficiently low or a cer- tain stop-criterion is reached. 45 & 8 & 8 Fracture Toughness 7 2 3 4 5 Case Number Fig. 2. K- prediction for straight crack. 316 ‘A. SEIBI and S. M. AL-ALAWL & Fracture Toughness: 8 8 case Number Fig. 3K. prediction for surface crack For the present work, the training process was performed using the NeuroShell® simulator. After several adjustments to the network parameters, the network converged to a threshold of 0.0001. The trained model prediction was in good agreement with the actual gains, hence, pro- ducing the R? value of 0.98 for K. for surface crack and 0.97 for K. for straight crack. These results show that approximately 98% and 97% of the variation in the dependent variables (output parameters) can be explained by the independent variables (input parameters) selected and the data set used. Having trained the network successfully, the next step is to test the trained network, using the test data, to judge its performance. Network testing and validation The generalization capability of the model was tested by presenting five test sets that were excluded from the experimental data prior to network training. Figures 2 and 3 provide the results and illustrate the relationship between actual and predicted fracture toughness values. To validate these results, three of the more common statistical parameters, generally used to determine model accuracy and performance, are used. These parameters are: mean absolute deviation (MAD), mean squared error (MSE), and mean absolute percent error (MAPE). MAD is the average of the absolute differences between the forecasted values and the observed or ex- perimental values. MSE is the average of the squared differences between forecasted and observed values, and MAPE is the absolute differences between the forecasted and observed values expressed as a percentage of the observed values. Statistical analysis of the results shown in Table 4 indicate that the R? value for the testing set was 0.998 for K. surface crack and 0.992 for K. straight crack. The MAPE was 9.47 and 9.25, respectively. These results demonstrate that the ANN-based model developed in this work can predict the fracture toughness K. with a high accuracy of approximately 91%. ANN AS AN ANALYTICAL TOOL An ANN can be a useful analytical tool in fracture mechanics. It can be used to determine the contribution of input variables and the variations in the output variable(s). It can also be used to analyse the effect of a specific input variable(s) such as crack geome- try, temperature, biaxiality, etc., on the model output variable(s). Once the model is developed and it produces accurate results, contribution of the different independent variables to the variation of the dependent variable values can be obtained from the model using the NeuroSheli® utility. Examination of the input variables’ contribution, as Table 4, Statistical analyses of the ANN model results ‘Statistical parameter ‘R, Surface crack , Straight crack ‘Mean absolute deviation (MAD) 2327 2.395 ‘Mean squared error (MSE) 10.893, 7588 ‘Mean absolute percent error (MAPE) 9.466 91246 Coefficient of determination (R?) 0.998 0992 Artificial neural networks 37 Table 5. The contribution of the input parameters to the output results Tapat iT T @ 2 ale « 8 parameter Contribution 55 650 35 60 83 4 103 % shown in Table 5, revealed that the temperature (7) had a substantial influence (65%) on the variation in fracture toughness. This result is followed by the effect of plate length to crack length (f) of approximately 10.3%. The contribution of crack depth to crack length ratio (a/c) and the crack length (2c) was 8.3% and 6.0%, respectively. Other factors such as the type of loading used (7), crack depth (a), and depth to thickness ratio (#) had only minor contri- butions to the variation in fracture toughness values. These contributions are shown in Table 5. Table 5 indicates that the average fracture toughness is mostly influenced by temperature as shown in Figs 4 and 5 for different crack depth-to-specimen thickness ratios, a, and Figs 6 and 7 for crack depth to crack length ratios, ajc, respectively. The general trend observed in these figures is that K- values increase with temperature up to an intermediate temperature and then fall down with small rises in between. The increase of K- up to the intermediate temperature is mainly related to the blunting process caused by the combined temperature-loading effects. On the other hand, for values higher than the intermediate temperature, less blunting takes place resulting in a decrease in K, values which is mainly governed by the yielding conditions. In ad- dition, both figures show that the fracture toughness values for straight cracks are lower than that for surface cracks for both loading conditions. Figures 4-7 are generated by inputting a/c values, that are not used in the experimental work, into the ANN model. The predictions for the average fracture toughness for both surface and straight cracks are obtained under different temperatures and loading conditions. It is also worth mentioning that the nature of loading does not have a significant effect on fracture toughness (less than 5.5%), indicating that the fracture toughness is independent of the uni- or biaxiality of loading especially that toward the crack onset, stress distribution is charac- terized by the triaxiality nature of the crack tip. CONCLUSIONS In this study, an ANN model has been developed and proved to be very efficient in predict- ing the fracture toughness of 7075-T651 aluminum alloy under uni- as well as biaxial loading conditions with a high degree of accuracy of 91%. The ANN model was also capable of deter- mining the contribution of each parameter that influences the values of the average fracture toughness. ‘As a result of using the ANN model in the analysis of the relationships between the average fracture toughness and crack geometry, type of loading and operating temperature, the follow- ing conclusions can be made: 45 Surface crack, Straight crack Practure toughness 3015100125180 17S 20025280275 300 Temperature Fig. 4. Effect of temperature on fracture toughness under uniaxial loading and changing « values. 318 ‘A, SEIBI and 8, M. AL-ALAWI 45 «ol Surface crack Fracture toughness S305 Too TSS 175 Baas 280 S300 ‘Temperature 8 8 8 Fracture Toughness 8 25 50 75 100 125 180 175 200 225 250 275 900 Temperature Fig. 6. Effect of temperature on fracture toughness under uniaxial loading and changing ac values. (1) Temperature has a pronounced effect on the average fracture toughness. (2) The fracture toughness depends on the crack geometry. (3) The biaxiality of the applied load has demonstrated little effect on the fracture toughness of the material. These special features and capabilities of ANN make it a good analytical tool that has potential use in the field of fracture mechanics. Further studies along this line are being carried out. Fracture Toughness Qo 8 2 80 75 100 125 160 178 200 225 20 275 300 Temperature Fig. 7. Effect of temperature on fracture toughness under biaxial loading and changing ae values. Astificial neural networks 319 REFERENCES Griffith, A. A., The phenomena of rupture and flow in solids. Transactions of the Royal Society of London, 1920, 221, 163-1982. Inwin, G. R., Handbuch der Physik, Vol. 6, Springer, Heidelberg, 1958, pp. 551-590. Irwin, G. R., Fracture mechanics. in Structural Mechanics. Pergamon Press, New York, 1960, pp. 557~592. Barsom, J. M., Development of the AASHTO fracture-toughness requirements for bridge steels, Engng Fracture Mech,, 1975, 73), 355-361. Barsom, J. M. and Rolfe, S. T., Fracture and Fatigue Controt in Structures, Applications of Fracture Mechanics, ‘Second Edition, Prentice Hall, Inc., 1987 Stanley, J., Introduction to Neural Networks, Third Edition, Sierra Madre, California Scientific Software, California, 1990. . Obermeir, K. K. and Barron, J. 3, Time to get fired up. BYTE Magazine, 1989, 14(8), 217-224. Lippmann, R. P., An introduction to computing with neural nets, /EEE ASSP Magazine, 1987, 4(2), 4-21 Kirkegaard, P. H. and Rytter, A., The use of neural networks for damage detection and location in a stee! member, Neural Networks and Combinatorial Optimization in Civil and Structural Engineering Conferences, Edinburgh, UK, 17-19 August 1993. Nakatsuji, T., Seki, S. and Kaku, T., Development of self-organizing traffic control system using neural network ‘model, Proceedings of the 4th International Conference on Microcomputers in Transportation, 1993, pp. 332-343. Islam, S. M. and Al-Alawi, S. M., Forecasting long-term electrical peak load and energy consumption for a fast growing utility using artificial neural networks, Proceedings of the International Power Engineering Conference, Singapore, 1995, pp. 690-695 AL-Alawi, S. M., Benjamin, C. O. and Omurtag, Y., Intelligent monitoring and control of large engineering projects, 6th Oklahoma Symposium on Artificial Inelligence, Tulsa, Oklahoma, 11-12 November 1992, pp. 115-124. Baghat, F. M., A fracture toughness estimation of plates containing surface cracks under certain conditions of tem- perature and loading, Ph.D thesis, The Pennsylvania State University, Pennsylvania, 1975, ‘Standard method of test for plane-srain fracture toughness of metalic materials, ASTM E399-12, 1973 Annual Book of ASTM Standards, Part 31, pp. 960-979. Hecht-Nielsen, R., Theory of the backpropagation neural network, Proceedings of the International Joint Conference (on Neural Networks, IEEE, Washington, D.C., 1989, 1, pp. 593-605, Lapedes, A. and Farber, R., How neural network works, Newral Information Processing Systems, American Institute of Physics, 1988, pp. 443-456. Rumelhart, D. E. and McClelland, J. L., Parallel distributed processing: exploration in the microstructure of cogni- tion, Foundations, Vol, 1, Cambridge, Massachusetts, MIT Press, 1986, Simpson, P. K., Artificial Neural Systems: Foundations, Paradigms, Applications, and Implementations, First Edition, Elmsford, New York, Pergamon Press Inc., 1990. Carpenter, W. C. and Hoffman, M. E., Training backpropagation neural networks. AI Expert, 1995, 10(3), 30-33. (Received 20 September 1995)

You might also like