The purpose of the paper is to model the relationship between the yield strength of austenitic stainless steel and its chemical composition through the employment of artificial neural network (ANN). The effect of annealing temperature is also taken into consideration.
The influence of network parameters, total number of neurons, number of neurons in a hidden layer, number of hidden layers and the interlayer distribution of neurons with a constant total neuron number, on the achievable training error is studied. Different learning rules available in MATLAB are used to assess the learning efficiencies of various networks.
It is found that increasing neuron number leads to a lowering of achievable training error up to a certain value beyond which training error remains constant. Increasing number of layers at constant total number of neurons causes a rise in the achievable training error. It is noted that if there is a sudden restriction of data flow in an ANN architecture, the achievable training error becomes higher. Interlayer distribution of neurons in ANNs used with different algorithms is found to have significant influence on the predictive performance of the networks.
From the study on metallurgical validation of the output of various ANNs, it appears that mere attainment of a lower training error is not sufficient to achieve better generalization. A convergent network topology is found to be better than a divergent one in respect of effectively describing the input‐output relationship in austenitic stainless steel.
