PSI - Issue 52

7

Vinit Vijay Deshpande et al. / Procedia Structural Integrity 52 (2024) 391–400 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

397

Fig. 7. Schematic architecture of both outer and inner neural networks.

4.2. Hybrid neural network Hybrid neural network utilizes the information learned by the inner network to better predict the response of the larger sized volume element. The training of this hybrid network is performed on data of larger sized volume element. The architecture of the network is shown in Fig. 8. The input layer is connected to inner neural network which is pretrained (in section 4.1). The output of this model along with the two macroscopic strains form an input to a single hidden layer that further connects to the last layer of output stresses. Through model selection, the number of neurons in this layer is fixed at 20 with the same L2 regularization. The same optimizer, activation functions and loss function as previous two models are used here.

Fig. 8. Schematic architecture of hybrid neural network. The root mean square errors of outer and hybrid neural networks on the predictions of stresses of the large volume element at = 45˚ are shown in Fig. 9 for different training datasets. Each network prediction is repeated 20 times for