Department of Pattern Recognition

(eng): The increasing complexity of modern constitutive models of cyclic metal plasticity requires more efficient ways to achieve their optimal calibration. Traditional approaches, such as random search combined with Nelder-Mead optimization, are computationally expensive. In addition, they struggle with highly non-convex functions that have numerous local minima and complex behavior, making these methods highly sensitive to initial conditions. While numerical refinement is key, a better prediction for its initial point directly saves costs. In this work, we focus only on the uniaxial cyclic loading, as it is the dominant part of a general calibration process for such a model and can also utilize a closed-form solution, further speeding up the procedure. We propose a neural network framework with a loss function that combines the loss on both the predicted parameters and the generated stress responses. This network is then used to predict an initial point for Nelder-Mead optimization. Our method was also compared to the non-gradient Tensor Train Optimization method on both synthetic data and measured experiments.