engleski

Deep learning-generated rod finite elements

Numerous technological innovations that provide clear benefits that society sorely needs require new and advanced materials with their superior properties. The unique mechanical behaviour of these materials is inadequately described by conventional models and therefore poorly understood. Consequently, machine learning approaches appear to be a potential tool for characterizing complex material properties. The stiffness matrix contains information on the mechanical properties of the material whose key component is the strain-displacement matrix. In this research, we utilized deep learning methods to generate a strain-displacement matrix at Gauss points. The intrinsic coordinate system-based isoparametric formulation was employed to derive the element stiffness matrix and equations. The proposed method is utilized to develop quadratic 1D linear elastic and functionally graded rod finite elements with varying Young’s modulus. The dataset included nodal coordinates, nodal displacements, and strain of the finite element. Several sets of data including training data, test data, and validation data were employed to obtain the best feasible prediction model. The numerical tests indicated a satisfying model's performance.

Deep learning, finite element, isoparametric formulation, functionally graded material

nije evidentirano