Development of a Neural Network Simulator for Studying the Constitutive Behavior of Structural Composite Materials

Table 3

Constitutive laws of 40% W-fiber reinforced BMG composite from various approaches. The parentheses in the inverse analysis represent standard deviation from five different runs. During the “forward optimization” process (3rd row), the most insensitive parameter (𝜃^W₁) was excluded from tuning, and the freezing temperature (𝑇) was also fixed at 355°C because it was found so experimentally from the previous study [15]. Compared to the manual fitting (1st row), the inverse ANN (2nd row) is much more efficient (faster), and forward ANN optimization (3rd row) is both more efficient and accurate.

	𝜎^W0 (MPa)	𝜎^W1 (MPa)	𝜃^W0 (MPa)	𝜃^W1 (MPa)	𝜎^BMG0 (MPa)	𝑛^BMG	𝑇 (°C)	MSE
Iterative manual FEM fitting	1305	725	120700	1330	1900	7.3	355	6.6𝐸−6
ANN inverse analysis	1234 (7)	729 (1)	566949 (666)	1335 (4)	1898 (1)	10.3 (0.1)	397 (3)	4.1𝐸−7
ANN forward optimization	1258	620	200000	1330	1900	8	355	1.5𝐸−7