Abstract

Casting is one of the most used processes to form metals like aluminium. A casting part can contain several defects that threaten its resistance. Shrinkage porosity is one of the major anomalies that designers try to avoid. For this purpose, rounds of numerical simulations should be performed with operating on a selection of parameters in order to minimize the presence of porosity in the casting part. In general, these approaches are time-cost with dependence on the complexity of the study case and the needed accuracy. In this paper, a methodology of data-driven porosity prediction for 3D parts is proposed in order to minimize the time-cost. A supervised learning algorithm is implemented to learn nodal porosity prediction using decision trees based method. A dataset is generated from a casting simulation software with operating on a selection of parameters. The training is realised on critical features vector extracted from nodal thermal history. Model order reduction method is used to interpolate thermal fields allover the parameter space. This interpolation is sufficiently accurate with minor errors. Promising results of shrinkage porosity prediction on a 3D study case are obtained. An evaluation of these results is performed with reference to the simulations results. This solution can contribute to open perspectives for more data-driven solutions that optimize the time-cost in the design stage.