Constitutive Modeling of Shape Memory Alloy Using Artificial Neural Network Konštutívne modelovanie zliatiny s tvarovou pamäťou s využitím umelej neurónovej siete K. P. Arulshri, K. P. Padmanaban, V. Selladurai Abstract This paper makes an attempt to develop constitutive model for Shape Memory Alloy (SMA) material behavior using Artificial Neural Network. The thermo mechanical behavior of SMA depends on many variables such as Prestrain, applied stress and temperature, these variables are also inter-dependent and characteristics of the SMA wire are studied by controlling one variable while allowing the other two variable to vary, determine the material coefficients. One dimensional behaviour of SMA is modeled using Brinson model which is based on energy balance equation with quantifiable engineering quantities and material coefficients. In this proposed model, Artificial Neural Network is used to learn complex non linear relationship in the modeling of constitutive behaviour of Shape Memory Alloy. This proposed model learn constitutive behaviour and store the knowledge in their connection weights. Conventional model is expressed only in terms of observable variable, but the proposed model has not only used observable variable but also the variable representing the internal behaviour. Due to the introduction of additional variable this type of model has larger degree of freedom in description and thus being superior to the conventional model. The Predicted results using Artificial Neural Network (ANN) is found to correlate well with experimental results when compared with the correlation between conventional models with experiments. Keywords: Constitutive modeling, SMA, ANN, Shape recovery stress Abstrakt Príspevok sa snaží rozvinúť konštutívny model správania sa materiálu pre zliatiny s tvarovou pamäťou (SMA) s využitím umelej neurónovej siete. Termo-mechanické vlastnosti SMA závisia na mnohých premenných ako napríklad predpätie, použitý tlak a teplota, tieto premenné sú taktiež závislé medzi sebou a charakteristika SMA je zoširoka študovaná ovládaním jednej z premenných, pokiaľ zvyšné dve sú premenlivé, určujúc materiálové koeficienty. Jednorozmerné chovanie sa SMA je namodelované s využitím Brinsonovho modelu, ktorý je postavený na rovnici energetickej rovnováhy s kvantifikovateľnými inžinierskymi množstvami a materiálovými koeficientmi. V navrhovanom modely je použitá umelá neurónová sieť pre pozorovanie nelineárneho vzťahu v modelovaní konštutívneho správania sa zliatiny s tvarovou pamäťou. Navrhnutý model pozoruje konštutívne správanie sa a uchováva tieto znalosti vo svojich vzťažných hodnotách. Konvenčný model je vyjadrený iba v zmysle sledovateľnej premennej, ale navrhnutý model nepoužíva iba sledovateľné premenné ale taktiež prezentujúce vnútorné vlastnosti. Kvôli zavedeniu ďalších premenných tento typ modelu má väčší počet stupňov voľnosti a preto je nadradený konvenčnému modelu. Očakávaný výsledok vyu- žívajúci umelú neurónovú sieť (ANN) pri porovnávaní korelácie medzi konvenčným modelom a experimentmi sa zisťuje ako dobre korelujúci s experimentálnymi výsledkami Kľúčové slová: konštutívne modelovanie, SMA, ANN, napätie pre obnovenie tvaru 1 Introduction Shape memory alloys (SMAs) are a class of materials processing very attractive mechanical properties. Under certain conditions they exhibit two unique effects, which are known in the literature as the shape memory effect and super elasticity. An initially deformed SMA can recover its predetermined low temperature shape during heating, demonstrating the shape memory effect, which can also be used to generate large internal recovery stresses. At higher temperatures super elastic behavior is observed and this effect is associated with large son-linear recovery strain during loading and unloading of the SMA. Azadi et al [1] developed a multi-dimensional continuum-level constitutive model of shape memory alloys exhibiting pseudo-elastic behavior. In their paper, the constitutive relation was constructed based on the gradient of a transformation potential function (effective stress), and the kinetics of transformation are expressed based on a set of transformation surfaces in stress-temperature space. Yuping Zhu et al [2] developed a constitutive model for magnetic shape memory alloys (MSMAs) through a combined consideration of micromechanical and thermodynamic theories. The kinetic equation was established in terms of the thermodynamic driving force derived from the reduction of Gibbs free energy of MSMA. The first discoveries in this field were reported in the literature a few decades ago. Chang and Read found the shape memory effect in an Au-Cd alloy. This direction of research has been successfully continued and many new alloys and materials, particularly polymers and ceramics, are still intensively investigated due to their shape memory properties. It can be appreciated from the summary above that many researchers have used various models of the SMA behaviour in order to investigate the static or dynamic performances of different composite material structures with SMA components. For this reason it is essential to address and emphasize the major differences and the similarities between the models available in the literature, in order to ensure the correctness of any numerical simulation based on such models. The main objective of the work reported in this paper is to provide results from a comparative study on the SMA behaviour. This study comprises results obtained from the use of three of the most popular one-dimensional models, as developed by Brinson [3] model with by neural network. Major the accuracy of the models is carefully examined against results obtained by the 61 authors from recent experimental measurements of the performance of a Ni-Ti SMA wire. Bodner et al [4] derived and applied the Constitutive equations for elastic-viscoplastic strain hardening materials. 2 Thermomechanical behavior of SMA The martensitic transformation is the basis characteristics of an SMA involved in all the unique characteristics of the SMAs, The martensitic transformation may be simply illustrated by the change in the martensitic volume fraction with respect to temperature and stress shown in figure.1. These thermomechanical behaviors have been described by a variety of constitutive models, and the differences between the actual behavior and the model have been presented previously. Therefore, this paper suggests the neural network model based on the experimental results. In this work, the suggested neural network model predicts the shape recovery force, which is the important factor to be well predicted in terms of control problem of SMA wires. ady, and it is fixed between two ends of the fixtures after measuring the length of the wire. Then the oven is set to a constant temperature. The load is gradually increased from zero state stress. Change in length of the wire can be taken from the LVDT; strain value can be calculated from that. After reaching a certain stress value (within elastic limit), unloading is done. Temperature vs stress plot at two differrent constant temperatures for three different samples. The modules increase with increase in temperature, due to phase transformation. The variation of stress with respect to temperature at lower strain is higher than at higher strain values. Linear segment of the stress-strain curve was used to determine the Young’s modulus of the wire. The procedure is repeated for different temperatures and for different samples. When the material has been transformed to the martensitic phase, it can be subjected to a significant plastic deformation. Such transformation requires a minimum load or stress value to deform it from its original shape when it is in Martensite phase, which varies as there is, a variation in temperature. The values of σ Scr , σ crf could be found by plotting transformation stress start and finish at constant temperature i.e. critical stress values are plotted with respect to the temperature which are obtained from the stress –strain plot at different constant temperatures. Fig. 1 Critical stresses for transformation of Martensite twin conversion as functions of temperature Obr. 1 Kritické napätia pre transformáciu martenzitického zrastenca ako funkcie teploty 3 Experimental details The tensometer setup, shown in figure 2, has two fixed end plated where two cylindrical bars that are fixed to the end plates. Two sliding bars with fixtures to hold the SMA wire are mounted on the cylindrical bars, where one end is fixed and other end is allowed to slide over the cylindrical rod. SMA wire is fixed between fixtures, between the fixed end and movable end. The movable rod is connected to a lead screw, which has mechanical rotary actuator controls loading rate. Lead screw arrangement is used to transfer the force from rotary actuator from movable end. As a tensile force acts on the wire when the movable end is pulled as the actuator is subjected to a clockwise rotation. Electrical resistive heating controls temperature of the SMA wire and load is measured using a load cell. The displacement of the wire is measured by LVDT. The core of the LVDT is mounted on the movable end, which indicates the displacement of the wire, and the body of the LVDT is fixed to the base of the setup. To avoid the influence of the surroundings, the wire should be kept in an enclosed chamber. Current supplied is measured and it is calibrated to the temperature values. Thermo-mechanical characteristics of SMAs are studied by keeping temperature as constant, and strain value is measured for the corresponding change in stress during loading and unloading. The temperature kept as constant during loading and unloading. Wire, which is to be tested, is heated above austenite finish temperature and cooled to obtain 100% Martensite and to remove the stresses developed alre- 62 Fig. 2 Tensometer Setup Obr. 2 Nastavenie tenzometra There is a linear relationship between phase transition temperature, and stress. The constants CM and C A the material properties, which describe the relationship of temperature and critical stress to induce transformation and it can be obtained by the rate of change of critical stress value with respect to change in temperature. The critical stress value below M S is taken to be constant. The setup has an oven where the SMA wire that is to be tested is fixed between two fixtures. The temperature control unit sets temperatures of the oven. The wire is fixed between fixtures, where the top end is fixed and the movable bottom end is attached to actuator loading system. Change in length of the wire can be obtained from the LVDT (Linear Variable Differential Transformer) Core of the LVDT is mounted on the wire, and the body is placed on the top of the oven. A load-cell and a thermocouple are used to measure the applied load, and to measure the temperature of the wire. Gradual temperature increase can be measured by the thermocouple, which is wound on the wire. Tightening of the wire increases the load, and correspondding change in length can be measured using LVDT as a sensor. The material properties of Ni-Ti alloy is given in Table 1. This setup is used to determine stress-strain curve at different temperatures. The SMA wire can also be loaded by adding the load on the base plate, at equal distance from the center of the wire. Oven temperature controls the temperature of the wire. The wire is kept in an enclosed chamber so as to avoid the environmental disturbances on the wire and heat transfer by, convection and radiation. The temperature of the chamber can be kept as constant while doing the experiment. The relations for the Young’s modulus and the phase transformation coefficient are the same as equations (2) and (3), respectively. In this model, a modified cosine model for the Martensite volume fraction is used and is divided into two parts. where ζS is the portion of the Martensite that is transformed due to applied stress and ζT is the portion of Martensite that is temperature induced. The equations also vary during transformation. For the (MÅA) transformation: For T>Ms and (3) σ scr + CM (T − M s ) < σ < σ crf + CM (T − M s ) ) ξT 0 (ξ S − ξ S 0 ) 1 − ξS 0 For T<Ms and σ Scr < σ < σ crf ξT = ξT 0 − ξS = ⎫⎪ 1 + ξ S 0 π 1− ξS0 ⎪⎧ cos ⎨ cr (σ − σ crf )⎬ + cr 2 2 ⎪⎭ ⎪⎩σ s − σ f ξT = ξT 0 − ξT 0 (ξ S − ξ S 0 ) + ΔTξ 1− ξS 0 For Mf<T<Ms and T<To 1 − ξT 0 [cos(aM (T − M f )) + 1] Δ Tξ = 2 end of transformation. These values are approximated from a stress-strain curve where the initial phase was 100% Martensite. On the curve, it is clear where transformation begins and ends so that these values are easily determined. For MÆA conversion the Martensite volume fraction is determined from the following relations: (9) For T>As and C A (T − A f ) < σ < C A (T − AS ) C M (T − M s ) ≤ σ eq ≤ C M (T − M f ) (10) (11) (12) 4 Evaluation of SMA characteristics 4.1 Isothermal force–displacement relation (2) ( start of transformation and σ crf is the critical stress at the Where all terms with a subscript ‘o’ denote initial condition and are equivalent to those defined in the Liang and Rogers model. Brinson [5] developed the third model. Brinson uses the same constitutive equation with some modifications. The modified relation is as follows: (1) σ − σ 0 = E (ξ )ε − E (ξ 0 )ε 0 + Ω(ξ )ξ 0 − Ω(ξ 0 )ξ S 0 + θ (T − T0 ) ⎧⎪ π ⎫⎪ 1 + ξ S 0 1− ξS 0 x cos⎨ cr σ − σ crf − CM (T − Ms) ⎬ + cr 2 2 ⎪⎩σ s − σ f ⎪⎭ In the above equations, σ Scr is the critical stress for the ξS = ξS 0 − 3.1 Brinson Model ξs = Δ Tξ = 0 ξS0 (ξ 0 − ξ ) ξ0 ξ ξT = ξT 0 − T 0 (ξ 0 − ξ ) ξS0 Table 1 Material properties of the Ni-Ti alloy Tab 1. Materiálové vlastnosti zliatiny Ni-Ti Property Value Martensite finish temperature Mf (˚C) 25 Martensite start temperature MS (˚C) 34 Austenite start temperature AS (˚C) 55 Austenite finish temperature AS (˚C) 80 Stress influence coefficient CM MPa ˚C-1 5 Stress influence coefficient CA MPa ˚C-1 11.5 Maximum residual strain εL 0.067 Young’s modulus EM (GPa) 15 Young’s modulus EA (GPa) 30 Coefficient of thermal expansion αM (˚C-1) 6.6 x 10-6 Coefficient of thermal expansion αA (˚C-1) 1.1 x 10-5 100 Critical stress σS (MPa) Critical stress σS (MPa) 170 ξ = ξ S + ξT Else Figure 3 shows the complete loading and unloading process of the SMA wire under various isothermal conditions. Brinson model the stress range in which the Martensite to austenite transformation may occur, can be expressed in equation (9) and the corresponding stress range for the reverse transformation is given by equation (1) and CA and CM are material constants related to the influence of the stress on the phase transformation, an increase of temperature increases the transformation stresses as in equations (9) and (10). Since no residual strain remains above 300C in figure 3, we can determine that Ms will be near 300C. A residual strain remains after the isothermal loading and unloading process decreases below about 300C in figure 3 and then heating above Af can restore the SMA wire with residual strain to the original shape. We can obtain CA, CM, Af , and Ms of equations (9) and (10) from the results of an isothermal strain-strain experiment as shown in figure 3, but the other two constants cannot be obtained because the stroke of the SMA wire is limited. Thus, when the shape recovery force is predicted in the following section, two constants will be properly assumed. (4) (5) (6) (7) (8) Fig. 3 Isothermal Stress-strain curve of an SMA wire at different temperature Obr. 3 Izotermická krivka napätie-posunutie drôtu SMA pri rozdielnych teplotách 63 4.2 Shape recovery force Figure 4 shows the result of the restrained recovery of an SMA wire at various initial deflections. Restraining a deformed SMA wire, while heating the wire above its transition temperature, will generate a large recovery force. This unique behavior can offer active movement to conventional catheters, which require motorless motion. The important feature, as an actuator, is that the greater the initial deformation of the SMA wire the larger the recovery force will be. These characteristics are also shown in figure 4. Based on the multidimensional constitutive relation, the onedimensional shear stress and strain relation of an SMA can be expressed as (1) where G is the elastic shear modulus, is the phase transformation tensor (which can be calculated from –DεL where D is the elastic modulus), εL is the maximum recoverable strain of an SMA, and is the thermoelastic tensor, which is related to the thermal expansion of SMA. T is the temperature and ξ is the internal variable, which describes the degree of the martensitic transformation. This parameter, referred to as the martensitic fraction, is defined as the ratio of martensite to the total volume of the SMA. In order to predict the experimental results using equation (4), the 12 following material constants are required: where DA and DM are Young’s moduli in austenite phase and martensite phase, respectively; ν is Poisson’s ratio, εL is the maximum recoverable strain, and ξ0 is the initial martensitic fraction. Fig. 4 Recovery stress against temperature at various initial stresses Obr. 4 Obnovovacie napätie verzus teplota pri rozdielnych počiatočných napätiach Calculated results using equation (4), with the assumptions of some material properties, do not show good agreement with the experimental data, as shown in figure 5. This conventional constitutive modeling has the following disadvantages. It is not practical because the constitutive modeling requires many experimental data to determine the 12 material constants. To obtain the shape recovery force, equation (4) should be solved by iteration because of nonlinearity of equation (4). Thus, it is a time consuming method. Such disadvantages can be improved by a new method of constitutive modeling using a neural network, which was originally proposed by Ghaboussi et al [6]. This methodology was then applied to constitutive modeling of concrete, composite material and current temperature of the SMA wire. One output neuron, which produces the shape recovery force, constituted the output layer of the neural network. Two hidden layers were made up of four neurons each. 64 Among the nine experimental data sets with different initial deflections from each other, three experimental data sets were used in training the neural network and the remainder. Fig. 5 Comparison of experimental data and calculated results of the SMA wire by conventional constitutive modeling Obr. 5 Porovnanie experimentálnych údajov a vypočítaných výsledkov drôtu SMA pri konvenčnom konštutívnom modelovaní Three experimental data sets, were compared with the predicted curves that resulted from the trained neural network. The training data are 24 points, which are randomly selected, in five experimental data sets. Figure 7 shows that the trained neural network predicts the shape recovery force very well. The developed neural network has two advantages compared with the conventional constitutive equations. Fig. 6 The architecture of an MLP for the prediction of the shape recovery stress Obr. 6 Architektúra MLP pre predvídanie napätia pre obnovu tvaru Firstly, the prediction of the shape recovery force is possible immediately through the real-time monitoring of the temperature of the SMA because the neural network method does not require a history of materials. However, if the conventional constitutive equations are used, it takes time to obtain the result because nonlinear materials such as SMA are path dependent. Secondly, the developed neural network method is so simple, since it requires only two input data. However, the conventional constitutive equation based on Brinson theory requires several input data, which may be impossible to obtain by conventional experiments. The application of the neural network is able to predict the almost exact shape recovery force with only two input variables. These two input parameters are the initial deflection and temperature of the SMA wire. The initial deflection is the most important design factor in the design of an active catheter and the temperature may be a feedback variable in the control of the SMA wire. Thus, the developed neural network method as shown in figure 6 can be applied to designing an active catheter and explaining the mathematical model of the SMA wire in a controller design. that the force feedback control is better than the temperature. K.P.Arulshri, Kongu Engg College, Perundurai, Erode, Tamil Nadu, India. S_p_arul@redifmail.com K.P.Padmanaban, PSNA College of Engineering & Technology, Dindigul, Tamilnadu, India. Land Line: +914512431050 padmarubhan@yahoo.co.in V.Selladurai Coimbatore Institute of Technology, Coimbatore, Tamil Nadu, India. References Fig. 7 Comparison of the experimental data and the predicted results of the SMA wire by the neural network Obr. 7 Porovnanie experimentálnych údajov a predvídaných výsledkov drôtu SMA neurónovou sieťou 5 Conclusion This paper was mainly aimed at describing the concept of the proposed model as the first step, thereby indicating its applicability only from simple experimental data, so various further studies are still left open. Depending on the type of experiment, one may try to describe material behaviors using various characteristic points in the experiment for the representation of internal variables. Strategies for selecting proper internal variables are thus important issues. An adequate technique for extracting training data from experimental data has to be proposed for each internal variable accordingly. The evaluation of the characteristics of an SMA wire actuator was performed in this work. The fabricated SMA wire shows a difference between the experimental results and the predicted results from the constitutive model by Brinson. Therefore, in this work, the combination of SMA modeling and neural network technology was first proposed. The predicted results by the trained neural network based on the experiments showed good agreement with the experimental results. In terms of the shape recovery stress, this technique requires only two input parameters (initial deflection and temperature), which can be important design parameters of the active catheter. The load- and temperature-following experiments were conducted in order to evaluate the feedback control characteristic when the feedback variables were assumed to be temperature and force in the active catheter. The experimental results showed that cooling and control methods were able to enhance the following characteristic of the SMA wire. It also informed us [1] Azadi, B. Rajapakse, R.K.N.D. and Maijer, D.M., Multidimensional constitutive modeling of SMA during unstable pseudoelastic behavior International Journal of Solids and Structures, Volume 44, Issue 20, 1 October 2007, Pages 6473-6490. [2] Yuping Zhu and Guansuo Dui, Micromechanical modeling of the stress-induced super-elastic strain in magnetic shape memory alloy, Mechanics of Materials, Volume 39, Issue 12, December 2007, Pages 1025-1034. [3] Brinson L C and Lammering R, Finite Element analysis of the behavior of Shape Memory Alloys and their Applications, International Journal of Solids and Structures, Vol.30, 1993,pp. 3261-3280. [4] Bodner S. R. and Y. Partom, ‘Constitutive equations for elastic-viscoplastic strain hardening materials’, Trans. ASME, J. Appl. Mech., 42, 385–389 (1975). [5] Brinson V, Review of Mechanics of Shape Memory Alloy Structures, Applied Mechanics Reviews, Vol.50, No.11, Part 1,November 1997,pp629-645. [6] Ghaboussi J, Garrett J H Jr andWu X 1991 Knowledgebased modeling of material behavior with neural networks J. Eng. echan. Division, ASCE 117 132–53. 65