artificial neural network model

Many efforts have been made to model the the hot deformation (dynamic recrystallization) flow curves of different materials. Phenomenological constitutive models, physical-based constitutive models and artificial neural network (ANN) models are the main methods used for this purpose. However, there is no report on the modeling of warm deformation (dynamic spheroidization) flow curves of any kind of steels. In this work, a neural network with feed forward topology and Bayesian regularization training algorithm was used to predict the warm deformation flow curves of a eutectoid steel. The experimental data was provided by sampling the dynamic spheroidization flow curves of the tested steel obtained from warm compression tests conducted over a temperature range of 620-770 °C with different strain rates in the range of 0.01-10 s-1. To develop the neural network model, the overal data was divided into three categries of training, validation and testing. The scatter diagrams together with the root mean square error (RMSE) criterion were used to evaluate the prediction performance of the developed model. The low calculated RMSE value of 4.15 MPa for the overall data showed the robustness of the developed ANN model in predicting the warm deformation flow curves of the tested steel. The results can be further used in the mathematical simulation of warm metal forming processes.

Bend existence causes changes in the flow pattern, velocity and the water surface profile. The ability to simulate three-dimensional flow pattern is an important and significant issues in curved channel. In the present study, using three-dimensional model of computational fluid dynamics (CFD) and artificial neural network (ANN) model of multi-Layer perceptron (MLP), two velocities and pressure variables on the channel bed with 90º sharp bend is predicted and compared. Also extensive experimental work has been conducted to measure the flow variables in this bend. Experimental results are used to train and test the neural network model accordingly. Comparison of the numerical with experimental results show that CFD model with average Root Mean Square Error (RMSE), 0.02 and 0.13 and ANN model with R2 (determination coefficient) value, 0.984 and 0.99 to predict velocity and pressure respectively, has reasonable accuracy. Also, velocity pattern and flow pressure with both numerical (CFD and ANN) models at any point of the field channel is predictable. Comparison of the CFD and ANN models show that the ANN model with the average value of Mean Absolute Error (MAE), 0.048 to CFD model with the average MAE, 0.06 in prediction of velocity and pressure has more accuracy. The present neural network with less time and cost in designing and implementation of curved channels than other expensive and time consuming experimental and computational models can be used.