The parameters of Powder Injection Molding (PIM) process were modeled by artificial neural networks (ANNs). The feed-forward multilayer perceptron was utilized and trained by back-propagation algorithm. Particle size, particle morphology, debinding time, and sintering temperature were taken into account and regarded as inputs of the ANN model. The outputs included relative density, wax loss, shrinkage, and hardness. The results obtained using the ANN model were in good agreement with the experimental data. In fact, they displayed an average R-value of 0.95 versus the experimental values. The optimum architecture of ANN was 7-4-1, in which the network was trained with Levenberg–Marquardt training algorithm. Thus, the ANN model can be used to evaluate, calculate, and forecast PIM process parameters.

Artificial neural networks (ANNs) as simplified model of mankind’s neural system, are capable of simulating and predicting real world complex problems which are challenging and expensive to model physically. In this study the correlation between the flow stresses and strain rate, temperature, strain in thermomechanical process of 40NICRMO8-4 alloy has been modelled. The results revealed that flow stress for every strain value is less at high temperatures compared to those at low temperatures and material resistance against deformation will also decrease as temperature goes down. Moreover, increasing in strain rate when temperature is constant results in recrystallization to happen in higher strain values at times shorter. The employed neural network for this study was a feed forward multilayer perceptron trained with common back propagation algorithm. Similar to any other ANNs, the employed network receives some parameters as inputs and delivers some as outputs. The inputs given to this model were temperature, strain and strain rate while flow stress parameter was collected as requested output. Outputs, with high precision of approximately 99% accuracy, were predicted and produced during training phase. Likewise, the predicted output of the ANN model achieved an R-value of about 0.99871 compared with of those experimental values. Best results were obtained with an ANN model consist of two hidden layers trained with Levenberg–Marquardt training algorithm.

No abstract