Comparative Study of Various Neural Network Types for Direct Inverse Material Parameter Identification in Numerical Simulations
Increasing product requirements in the mechanical engineering industry and efforts to reduce time-to-market demand highly accurate and resource-efficient finite element simulations. The required parameter calibration of the material models is becoming increasingly challenging with regard to the growing variety of available materials. Besides the classical iterative optimization-based parameter identification method, novel machine learning-based methods represent promising alternatives, especially in terms of efficiency. However, the machine learning algorithms, architectures, and settings significantly affect the resulting accuracy. This work presents a comparative study of different machine learning algorithms based on virtual datasets with varying settings for the direct inverse material parameter identification method. Multilayer perceptrons, convolutional neural networks, and Bayesian neural networks are compared; and their resulting prediction accuracies are investigated. Furthermore, advantages in material parameter identification by uncertainty quantification using the Bayesian probabilistic approach are examined and discussed. The results show increased prediction quality when using convolutional neural networks instead of multilayer perceptrons. The assessment of the aleatoric and epistemic uncertainties when using Bayesian neural networks also demonstrated advantages in evaluating the reliability of the predicted material parameters and their influences on the subsequent finite element simulations.