This proposal proposes a surrogate model approach to obtain reliable simulations for costly expensive additive manufacturing simulations (such as microstructure/thermomechanical modeling).

This proposal proposes a surrogate model approach to obtain reliable simulations for costly expensive additive manufacturing simulations (such as microstructure/thermomechanical modeling). The method is beneficial for AM simulation, where we have the repetitive local/small-scale finite element simulations. The small-scale can be substituted by a surrogate model. Enough data through finite element simulation should be generated for training the algorithm. Compared to other machine learning methods, a surrogate model generates a metamodel by analyzing the results of a small number of simulations. Surrogate models are used to substitute expensive simulations with cheap ones to simulate computational models. These models use computational algorithms and intend to be substituted as black boxes. The surrogate model does not need information about the corresponding constitutive equations and works in a data-driven way, which uses sets of input and output data points of the simulation. The approach employs a few sets of expensive simulations in an intelligent way to construct a new model that can reproduce the behavior of the original simulation. The metamodel utilizes polynomial chaos expansion to model the dependence between reduced input and output data. Polynomial chaos expansion (PCE) was developed to define the evolution of uncertainties in systems with probabilistic input. It reproduces the behavior of the original model with a set of orthogonal polynomial functions. It is worth mentioning that for conditions where the associated uncertainty is unacceptable, additional local simulations can be devised and joined the training data to decrease the uncertainty level. Moreover, the proposed surrogate model can use principal component analysis to reduce the dimensions of input and output space to further lower the computational cost. The aim is to reduce the computational cost for AM finite element simulation. By manipulating a particular machine learning approach, the computational cost of repetitive simulations can be decreased. A systematic set of experiments should be designed to identify the sensitivity of the output parameters to the input boundary conditions for small-scale simulations. The reduction is obtained by replacing the small-scale finite element simulations with a surrogate model. Results from these simulations can be utilized to train a cheap surrogate model without significantly decreasing the efficacy. This method leads to a notable cost reduction for AM simulations.