Logistic Sigmoid Transfer Function Research Articles

The prediction of maximum temperature for single chips’ cooling using artificial neural networks

A CFD simulation usually requires extensive computer storage and lengthy computational time. The application of artificial neural network models to thermal management of chips is still limited. In this study, the main objective is to find a neural network solution for obtaining suitable thickness levels and material for a chip subjected to a constant heat power. To achieve this aim a neural network is trained and tested using the results of the CFD program package Fluent. The back-propagation learning algorithm with three different variants, single layer and logistic sigmoid transfer function is employed in the network. By using the weights of the network, various formulations are designed for the output. The network has resulted in R2 values of 0.999, and the mean% errors smaller than 0.8 and 0.7 for the training and test data, respectively. The analysis is extended for different thickness and input power values. Comparison of some randomly selected results obtained by the neural network model and the CFD program has yielded a maximum error of 1.8%, mean absolute percentage error of 0.55% and R2 of 0.99994.

Estimation of solar radiation over Turkey using artificial neural network and satellite data

Abstract This study introduces artificial neural networks (ANNs) for the estimation of solar radiation in Turkey (26–45 E and 36–42 N). Resilient propagation (RP), Scale conjugate gradient (SCG) learning algorithms and logistic sigmoid transfer function were used in the network. In order to train the neural network, meteorological data for the period from August 1997 to December 1997 for 12 cities (Antalya, Artvin, Edirne, Kayseri, Kutahya, Van, Adana, Ankara, Istanbul, Samsun, Izmir, Diyarbakir) spread over Turkey were used as training (nine stations) and testing (three stations) data. Meteorological and geographical data (latitude, longitude, altitude, month, mean diffuse radiation and mean beam radiation) are used in the input layer of the network. Solar radiation is the output. However, solar radiation has been estimated as monthly mean daily sum by using Meteosat-6 satellite C3 D data in the visible range over 12 cities in Turkey. Digital counts of satellite data were converted into radiances and these are used to calculate the albedos. Using the albedo, the cloud cover index of each pixel was constructed. Diffuse and direct component of horizontal irradiation were calculated as a function of optical air mass, turbidity factor and Rayleigh optical thickness for clear-sky. Using the relation between clear-sky index and cloud cover index, the solar irradiance for any pixel is calculated for Physical method. RMS between the estimated and ground values for monthly mean daily sum with ANN and Physical method values have been found as 2.32 MJ m −2 (54 W/m 2 ) and 2.75 MJ m −2 (64 W/m 2 ) (training cities), 3.94 MJ m −2 (91 W/m 2 ) and 5.37 MJ m −2 (125 W/m 2 ) (testing cities), respectively.

Modelling of residual stresses in the shot peened material C-1020 by artificial neural network

This study consists of two cases: (i) The experimental analysis: Shot peening is a method to improve the resistance of metal pieces to fatigue by creating regions of residual stress. In this study, the residual stresses induced in steel specimen type C-1020 by applying various strengths of shot peening, are investigated using the electrochemical layer removal method. The best result is obtained using 0.26 mm A peening strength and the stress encountered in the shot peened material is −276 MPa, while the maximum residual stress obtained is −363 MPa at a peening strength of 0.43 mm A. (ii) The mathematical modelling analysis: The use of ANN has been proposed to determine the residual stresses based on various strengths of shot peening using results of experimental analysis. The back-propagation learning algorithm with two different variants and logistic sigmoid transfer function were used in the network. In order to train the neural network, limited experimental measurements were used as training and test data. The best fitting training data set was obtained with four neurons in the hidden layer, which made it possible to predict residual stress with accuracy at least as good as that of the experimental error, over the whole experimental range. After training, it was found the R 2 values are 0.996112 and 0.99896 for annealed before peening and shot peened only, respectively. Similarly, these values for testing data are 0.995858 and 0.999143, respectively. As seen from the results of mathematical modelling, the calculated residual stresses are obviously within acceptable uncertainties.

The experimental investigation of the effects of uncoated, PVD- and CVD-coated cemented carbide inserts and cutting parameters on surface roughness in CNC turning and its prediction using artificial neural networks

In this study the machining of AISI 1030 steel (i.e. orthogonal cutting) uncoated, PVD- and CVD-coated cemented carbide insert with different feed rates of 0.25, 0.30, 0.35, 0.40 and 0.45 mm/rev with the cutting speeds of 100, 200 and 300 m/min by keeping depth of cuts constant (i.e. 2 mm), without using cooling liquids has been accomplished. The surface roughness effects of coating method, coating material, cutting speed and feed rate on the workpiece have been investigated. Among the cutting tools—with 200 mm/min cutting speed and 0.25 mm/rev feed rate—the TiN coated with PVD method has provided 2.16 μm, TiAlN coated with PVD method has provided 2.3 μm, AlTiN coated with PVD method has provided 2.46 μm surface roughness values, respectively. While the uncoated cutting tool with the cutting speed of 100 m/min and 0.25 mm/rev feed rate has yielded the surface roughness value of 2.45 μm. Afterwards, these experimental studies were executed on artificial neural networks (ANN). The training and test data of the ANNs have been prepared using experimental patterns for the surface roughness. In the input layer of the ANNs, the coating tools, feed rate ( f) and cutting speed ( V) values are used while at the output layer the surface roughness values are used. They are used to train and test multilayered, hierarchically connected and directed networks with varying numbers of the hidden layers using back-propagation scaled conjugate gradient (SCG) and Levenberg–Marquardt (LM) algorithms with the logistic sigmoid transfer function. The experimental values and ANN predictions are compared by statistical error analyzing methods. It is shown that the SCG model with nine neurons in the hidden layer has produced absolute fraction of variance ( R 2) values about 0.99985 for the training data, and 0.99983 for the test data; root mean square error ( RMSE) values are smaller than 0.00265; and mean error percentage ( MEP) are about 1.13458 and 1.88698 for the training and test data, respectively. Therefore, the surface roughness value has been determined by the ANN with an acceptable accuracy.

Identification of Stored-Grain Age Using Electronic Nose by Ann

In this study identification of electronic nose for wheat age was conducted. An electronic nose was employed to collect and classify aroma signals from different age of wheat. A back-propagation artificial neural network (ANN) with a tan-sigmoid transfer function in the hidden layers and a log-sigmoid transfer function in the output layer was employed and trained with extracted features from the data collected by electronic nose to identify the wheat age. This preliminary investigation succeeded in identifying five different age of wheat in the laboratory. Artificial neural networks (ANNs) were trained with average features from the data collected by the electronic nose to identify age of wheat, the result showed best performance at same hidden neurons for training and testing. The percentage correct identification rate of ANN was: 92%, 98%, and 94%. The train epochs were 1012, 397, and 583. The networks with average features have satisfactory classification accuracy for both the training and the test set.

Determination of efficiency of flat-plate solar collectors using neural network approach

In this study, a new formula based on artificial neural network (ANN) technique was developed to determine the efficiency of flat-plate solar collectors. In practice, the ANN model can be used for modeling the efficiency of solar collectors with complex structures when other models may have difficulties. Logistic sigmoid transfer function was used in the network. Meteorological data of summer session (from July to September) for Ankara were used as training data in order to train the neural network. Surface temperature in collector, date, time, solar radiation, declination angle, azimuth angle and tilt angle are used in the input layer of the network. The efficiency of flat-plate solar collector is in the output layer. The results show that the maximum and minimum deviations were found 2.558484 and 0.001969, respectively. The advantages of ANN model compared to the conventional testing methods are speed, simplicity, and the capacity of the ANN to learn from examples.

The Use of Artificial Neural Networks in Simulation of Mobile Ground Vehicles

In this study, we have developed a platform which incorporates Artificial Neural Networks (ANNs) in simulating body dynamics of mobile ground vehicles (e.g. cars). This is a part of our research project in which we plan to provide a platform for educating the driver candidates in virtual environments: where the drivers can be educated fully in “Artificial Cities”. To start with, 6 different makes of cars with different engine properties has been simulated with the appropriate data provided by the manufacturers and rules of physics. A joystick steering wheel has been used to produce the necessary inputs for the ANN based physics engine. To train the network, Scaled Conjugate Gradient (SCG) and Levenberg-Marquardt (LM) learning algorithms and a logistic sigmoid transfer function have been used. The statistical error levels are negligible. The Absolute Fraction of Variance (R2) values for both the training and test data are about 99.999% and the mean error value for both data group is lesser than 0.5%.

A Generalized Neural Network Model of Refrigerant Mass Flow Through Adiabatic Capillary Tubes and Short Tube Orifices

Adiabatic capillary tubes and short tube orifices are widely used as expansive devices in refrigeration, residential air conditioners, and heat pumps. In this paper, a generalized neural network has been developed to predict the mass flow rate through adiabatic capillary tubes and short tube orifices. The input/output parameters of the neural network are dimensionless and derived from the homogeneous equilibrium flow model. Three-layer backpropagation (BP) neural network is selected as a universal function approximator. Log sigmoid and pure linear transfer functions are used in the hidden layer and the output layer, respectively. The experimental data of R12, R22, R134a, R404A, R407C, R410A, and R600a from the open literature covering capillary and short tube geometries, subcooled and two-phase inlet conditions, are collected for the BP network training and testing. Compared with experimental data, the overall average and standard deviations of the proposed neural network are 0.75% and 8.27% of the measured mass flow rates, respectively.

Prospects for Future Projections of the Basic Energy Sources in Turkey

The main goal of this study is to develop the energy sources estimation equations in order to estimate the future projections and make correct investments in Turkey using artificial neural network (ANN) approach. It is also expected that this study will be helpful in demonstrating energy situation of Turkey in amount of EU countries. Basic energy indicators such as population, gross generation, installed capacity, net energy consumption, import, export are used in input layer of ANN. Basic energy sources such as coal, lignite, fuel-oil, natural gas and hydraulic are in output layer. Data from 1975 to 2003 are used to train. Three years (1981, 1994 and 2003) are only used as test data to confirm this method. Also, in this study, the best approach was investigated for each energy sources by using different learning algorithms (scaled conjugate gradient [SCG] and Levenberg-Marquardt [LM]) and a logistic sigmoid transfer function in the ANN with developed software. The statistical coefficients of multiple determinations (R2-value) for training data are equal to 0.99802, 0.99918, 0.997134, 0.998831 and 0.995681 for natural gas, lignite, coal, hydraulic, and fuel-oil, respectively. Similarly, these values for testing data are equal to 0.995623, 0.999456, 0.998545, 0.999236, and 0.99002. The best approach was found for lignite by SCG algorithm with seven neurons so mean absolute percentage error (MAPE) is equal to 1.646753 for lignite. According to the results, the future projections of energy indicators using ANN technique have been obviously predicted within acceptable errors. Apart from reducing the whole time required, the importance of the ANN approach is possible to find solutions that make energy applications more viable and thus more attractive to potential users.

Prediction of Extrudate Properties Using Artificial Neural Networks

A backpropagation artificial neural network (ANN) model was developed to predict the properties of extrudates generated by extrusion cooking of fish muscle-rice flour blend in a single screw extruder. Experimental data obtained in a previous study on extrudate properties of expansion ratio, bulk density and hardness at different combinations of operating variables of barrel temperature, feed content and feed moisture had been analysed using response surface methodology (RSM). A backpropagation neural network model was implemented in MATLAB and was trained for operating variables (inputs) and for each individual measured extrudate properties expansion ratio ER, bulk density BD and harndess H (outputs). The optimized network indicated that one hidden layer with a learning rate of 0.1, steep descent learning rule, 100 000 epochs and a logistic sigmoid transfer function predicted the extrudate properties better than RSM. The agreement of the ANN model with the experimental values, expressed as sum of squared error values, was 9.8 × 10 −7 for ER, 5.8 × 10 −2 for BD and 3.8 × 10 −3 for H. The ANN prediction for the optimized process conditions was superior to the RSM values, with percentage errors of +6.06% (ER), +4.08% (BD) and −14.28% (H).

Application of artificial neural networks for the wind speed prediction of target station using reference stations data

In this study, artificial neural networks (ANNs) were applied to predict the mean monthly wind speed of any target station using the mean monthly wind speeds of neighboring stations which are indicated as reference stations. Hourly wind speed data, collected by the Turkish State Meteorological Service (TSMS) at 8 measuring stations located in the eastern Mediterranean region of Turkey were used. The long-term wind data, containing hourly wind speeds, directions and related information, cover the period between 1992 and 2001. These data were divided into two sections. According to the correlation coefficients, reference and target stations were defined. The mean monthly wind speeds of reference stations were used and also corresponding months were specified in the input layer of the network. On the other hand, the mean monthly wind speed of the target station was utilized in the output layer of the network. Resilient propagation (RP) learning algorithm was applied in the present simulation. The hidden layers and output layer of the network consist of logistic sigmoid transfer function (logsig) and linear transfer function (purelin) as an activation function. Finally, the values determined by ANN model were compared with the actual data. The maximum mean absolute percentage error was found to be 14.13% for Antakya meteorological station and the best result was found to be 4.49% for Mersin meteorological station.

Experimental and Theoretical Investigations of Mouldability for Feedstocks Used in Powder Injection Moulding

Experimental and theoretical analyses of mouldability for feedstocks used in powder injection moulding are performed. This study covers two main analyses. (i)The experimental analysis: the barrel temperature, injection pressure, and flow rate are factors for powder injection moulding (PIM). Powder-binder mixture used as feedstock in PIM requires a little more attention and sensitivity. Obtaining the balance among pressure, temperature, and especially flow rate is the most important aspect of undesirable conclusions such as powder-binder separation, sink marks, and cracks in moulded party structure. In this study, available feedstocks used in PIM were injected in three different cavities which consist of zigzag form, constant cross-section, and stair form (in five different thicknesses) and their mouldability is measured. Because of the difference between material and binder, measured lengths were different. These were measured as 533 mm, 268 mm, 211 mm, and 150 mm in advanced materials trade marks Fe–2Ni, BASF firm Catamould A0-F, FN02, and 316L stainless steel, respectively. (ii)The theoretical analysis: the use of artificial neural network (ANN) has been proposed to determine the mouldability for feedstocks used in powder injection moulding using results of experimental analysis. The back-propagation learning algorithm with two different variants and logistic sigmoid transfer function were used in the network. In order to train the neural network, limited experimental measurements were used as training and test data. The best fitting training data set was obtained with three and four neurons in the hidden layer, which made it possible to predict yield length with accuracy at least as good as that of the experimental error, over the whole experimental range. After training, it was found that theR2values are 0.999463, 0.999445, 0.999574, and 0.999593 for Fe–2Ni, BASF firm Catamould A0-F, FN02, and 316L stainless steel, respectively. Similarly, these values for testing data are 0.999129, 0.999666, 0.998612, and 0.997512, respectively. As seen from the results of mathematical modeling, the calculated yield lengths are obviously within acceptable uncertainties.

Determination of residual stresses based on heat treatment conditions and densities on a hybrid (FLN2-4405) powder metallurgy steel using artificial neural network

This paper presents a new approach based on artificial neural networks (ANNs) to determine the residual stresses in PM steel based nickel (FLN2-4405). This study consists of two cases: (i) The experimental analysis: The measurements of residual stresses were carried out by electrochemical layer removal technique. The values and distributions of residual stresses occurring in PM steel processed under various densities (6.8, 7.05, 7.2 and 7.4 g/cm 3) and heat treatment conditions (sintered at 2050 °F, sintered at 2300 °F, quenching-tempered, and sinter-hardened) were determined. In most of the experiments, tensile residual stresses were recorded in surface of samples. The residual stress distribution on the surface of the PM steels is affected by the heat treatment conditions and density. Maximum values of residual stresses on the surface were observed sinter hardened condition and 7.4 g/cm 3 density. (ii) The mathematical modeling analysis: The use of ANN has been proposed to determine the residual stresses based on heat treatment conditions and densities in PM steel using results of experimental analysis. The back propagation learning algorithm with two different variants and logistic sigmoid transfer function were used in the network. In order to train the neural network, limited experimental measurements were used as training and test data. The best fitting training data set was obtained with four and five neurons in the hidden layer, which made it possible to predict residual stress with accuracy at least as good as that of the experimental error, over the whole experimental range. After training, it was found the R 2 values are 0.999244, 0.999025, 0.999664 and 0.999322 for sintered at 2050 °F, sintered at 2300 °F, quenching-tempered, and sinter-hardened, respectively. Similarly, these values for testing data are 0.998354, 0.99706, 0.999607 and 0.999205, respectively. As seen from the results of mathematical modeling, the calculated residual stresses are obviously within acceptable uncertainties.

Comparative study of mathematical and experimental analysis of spark ignition engine performance used ethanol–gasoline blend fuel

This study consists of two cases: (i) The experimental analysis: Ethanol obtained from biomass can be used as a fuel in spark ignition engines. As renewable energy source ethanol, due to the high octane number, low emissions and high engine performance is preferred alternative fuel. First stage of this study, ethanol–unleaded gasoline blends (E10, E20, E40 and E60) were tested in a single cylinder, four-stroke spark ignition and fuel injection engine. The tests were performed by varying the ignition timing, relative air–fuel ratio (RAFR) and compression ratio at a constant speed of 2000 rpm and at wide open throttle (WOT). Effect of ethanol–unleaded gasoline blends and tests variables on engine torque and specific fuel consumption were examined experimentally. (ii) The mathematical modeling analysis: The use of ANN has been proposed to determine the engine torque and specific fuel consumption based on the ignition timing, RAFR and compression ratio at a constant speed of 2000 rpm and at WOT for different fuel densities using results of experimental analysis. The back-propagation learning algorithm with two different variants and logistic sigmoid transfer function were used in the network. In order to train the neural network, limited experimental measurements were used as training and test data. The best fitting training data set was obtained Levenberg–Marquardt (LM) algorithm with five neurons in the hidden layer, which made it possible to the engine torque and specific fuel consumption with accuracy at least as good as that of the experimental error, over the whole experimental range. After training, it was found the R 2 values are 0.999996 and 0.999991 for, the engine torque and specific fuel consumption, respectively. Similarly, these values for testing data are 0.999977 and 0.999915, respectively. As seen from the results of mathematical modeling, the calculated engine torque and specific fuel consumption are obviously within acceptable uncertainties.

Calculation for the thermodynamic properties of an alternative refrigerant (R508b) using artificial neural network

This study proposes a alternative approach based on artificial neural networks (ANNs) to determine the thermodynamic properties – specific volume, enthalpy and entropy – of an alternative refrigerant (R508b) for both saturated liquid–vapor region (wet vapor) and superheated vapor region. In the ANN, the back-propagation learning algorithm with two different variants, namely scaled conjugate gradient (SCG) and Levenberg–Marquardt (LM), and Logistic Sigmoid transfer function were used to determine the best approach. The most suitable algorithm and with appropriate number of neurons (i.e. 7) in the hidden layer is found to be the LM algorithm which has provided the minimum error. For wet vapor region, R 2 values – which are errors known as absolute fraction of variance – are 0.983495, 0.969027, 0.999984, 0.999963, 0.999981, and 0.999975, for specific volume, enthalpy and entropy for training and testing, respectively. Similarly, for superheated vapor, they are: 0.995346, 0.996947, 0.999996, 0.999997, 0.999974, and 0.999975, for training and testing, respectively. According to the regression analysis results, R 2 values are 0.9312, 0.9708, 0.9428, 0.9343, 0.967 and 0.9546 for specific volume, enthalpy and entropy for wet vapor region and superheated vapor, respectively. The comparisons of the results suggest that, ANN provided results comfortably within the acceptable range. This study, deals with the potential application of the ANNs to represent PVTx (pressure–specific volume–temperature–vapor quality) data. Therefore, reducing the risk of experimental uncertainties and also removing the need for complex analytic equations requiring long computational time and efforts.

Forecasting Net Energy Consumption Using Artificial Neural Network

The main goal of this study is to develop the equations for forecasting net energy consumption (NEC) using the artificial neural network (ANN) technique in order to determine the future level of the energy consumption in Turkey. Logistic sigmoid transfer function was used in the network. In order to train the neural network, population, and gross generation, installed capacity and years is used in input layer of network. The net energy consumption is in output layer. The input values in 1965, 1981, and 1997 are only used as test data to confirm this method. The statistical coefficient of multiple determinations (R 2-value) is equal to 0.9999 and 1 for training and test data, respectively. According to the results, the NEC using the ANN technique has been obviously predicted within acceptable errors. Apart from reducing the whole time required, the importance of the ANN approach is possible to find solutions that make energy applications more viable and thus more attractive to potential users. It is also expected that this study will be helpful in developing highly applicable and productive planning for energy policies.

Predictions of temperature distributions on layered metal plates using artificial neural networks

The temperature distribution influences the amount of energy needed to heat a body. The benefits of using multi-layered metal plates (MMP) are due to the requirement of a regular temperature distribution on the opposite side with one side heated irregularly. The factors that affect the regular distribution of the temperature in such a structure are the thickness of the layers and the materials themselves, since for different materials, heat conduction coefficients, density and specific heat values change. In this study, the main objective is to find a neural network solution for the problem of the non-regular distribution of temperature on the non-heated side of an irregularly heated MMP consisting of two layers of Cu/CrNi and Al/CrNi in order to obtain the optimum thickness levels for the layers. To achieve this aim, the results of the finite elements method (FEM) produced by the program package ANSYS have been used to train and test the network. They are the coefficient of heat conduction ( K ), specific heat ( C ), density ( D ), temperature ( T ) and layer thickness ( L ), which are used as the input layer, while the outputs are the maximum, minimum and mean temperature values of the materials. The back propagation learning algorithm with three different variants, single layer and logistic sigmoid transfer function have been used in the network. By using the weights of the network, formulations have been given for each output. The network has yielded R 2 values of 0.999 and the mean percent errors are smaller than 0.8 for the training data, while the R 2 values are about 0.999 and the mean percent errors are smaller than 0.7 for the test data. The analysis has been extended for different materials and for the different temperature values that have been applied. The Al/CrNi laminated plate has a lower temperature gradient distribution on the upper (or non-heated) surface due to its lesser heat conductivity compared to the Cu/CrNi steel. The thickness of 8 mm provides the best results among the alloys that have been considered.

Modelling of yield length in the mould of commercial plastics using artificial neural networks

According to the yield length of a plastic, whether a mould is filled fully or not can be estimated. In this study, a new formula based on various injection parameters was developed to determine the yield length in plastic moulding of the commercial plastics, the most widely used ones are low-density polyethylene, high-density polyethylene, polystyrene and polypropylene, by artificial neural network (ANN). This study covers two main objectives: (i) the yield properties of plastics have been investigated at various injection parameters (cylinder temperature, injection pressure, injection flow rate and mould temperature) in a mould including spiral chutes as experimental; (ii) the yield length of the commercial plastics based on various measured injection parameters (cylinder temperature, injection pressure, injection flow rate and mould temperature) in a mould was described by ANN using experimental data. Some experimental data were used as test data, and these values were not used for training. Scaled conjugate gradient and Levenberg–Marquardt (LM) learning algorithms and logistic sigmoid transfer function were used in the network. The results show that the maximum mean absolute percentage error (MAPE) was found to be 1.574969% and R 2 values 0.99964. The best approach was found in the LM algorithm with six neurons. The MAPE and R 2 for testing data were 1.849 and 0.9995, respectively, similar to algorithm and neurons. This study is considered to be helpful in predicting the yield length in the mould whose injection parameters are known. The present method greatly reduces the time required by design engineers to find the optimum solution and in many cases reaches a solution that could not be easily obtained from simple modeling programs.

Performance and exhaust emissions of a biodiesel engine

In this study, the applicabilities of Artificial Neural Networks (ANNs) have been investigated for the performance and exhaust-emission values of a diesel engine fueled with biodiesels from different feedstocks and petroleum diesel fuels. The engine performance and emissions characteristics of two different petroleum diesel-fuels (No. 1 and No. 2), biodiesels (from soybean oil and yellow grease), and their 20% blends with No. 2 diesel fuel were used as experimental results. The fuels were tested at full load (100%) at 1400-rpm engine speed, where the engine torque was 257.6 Nm. To train the network, the average molecular weight, net heat of combustion, specific gravity, kinematic viscosity, C/H ratio and cetane number of each fuel are used as the input layer, while outputs are the brake specific fuel-consumption, exhaust temperature, and exhaust emissions. The back-propagation learning algorithm with three different variants, single layer, and logistic sigmoid transfer function were used in the network. By using weights in the network, formulations have been given for each output. The network has yielded R 2 values of 0.99 and the mean % errors are smaller than 4.2 for the training data, while the R 2 values are about 0.99 and the mean % errors are smaller than 5.5 for the test data. The performance and exhaust emissions from a diesel engine, using biodiesel blends with No. 2 diesel fuel up to 20%, have been predicted using the ANN model.

Effective parameters modeling in compression of an austenitic stainless steel using artificial neural network

In this study, the prediction of flow stress in 304 stainless steel using artificial neural networks (ANN) has been investigated. Experimental data earlier deduced—by [S. Venugopal et al., Optimization of cold and warm workability in 304 stainless steel using instability maps, Metall. Trans. A 27A (1996) 126–199]—were collected to obtain training and test data. Temperature, strain-rate and strain were used as input layer, while the output was flow stress. The back propagation learning algorithm with three different variants and logistic sigmoid transfer function were used in the network. The results of this investigation shows that the R 2 values for the test and training data set are about 0.9791 and 0.9871, respectively, and the smallest mean absolute error is 14.235. With these results, we believe that the ANN can be used for prediction of flow stress as an accurate method in 304 stainless steel.