On the efficient texture representation in neural network methods for plastic deformation modelling

Representation of crystallographic texture and its evolution is critical for efficient mesoscale modelling of plastic deformation. Although Euler angles are widely adopted to represent texture, the high dimensionality and unclear correlation with stress responses compromise their effectiveness in neural network-based deformation modelling, especially when applied to diverse textures. In the current work, a texture component description method is adopted in the neural network developments to predict the texture evolution and its effects by leveraging the Fourier coefficients of generalized spherical harmonics functions, where a complex texture can be described through a linear combination of base components’ coefficients. The results with Random, Cube and S texture components illustrate that the component-based description strategy with Fourier coefficients enables the model to predict the responses of various texture component combinations without complex training and fine-tuning processes. The novel description is argued to be an efficient texture representation method for neural network models.