Design optimization poses significant challenges due to the substantial expensive and time-consuming characteristic of simulations. To counteract this issue, deep learning-based surrogate models have recently emerged as an effective solution. However, current research has primarily focused on applying these models to Computational Fluid Dynamics, with limited studies in the area of Linear Elasticity. Unlike previous work, the investigated structures are 3D with notable variations in the simulation domains, resulting in significant changes to the learning domain of the network. The proposed Convolutional Neural Network is based on the U-Net architecture and is trained in a supervised manner to learn the displacement and stress results. The problem is treated as an image-to-image learning task, with the simulation domain encoded through a binary mask. The input images are associated with the spatial distribution of material parameters, the applied boundary conditions, and the force density distribution. The accuracy of the model has been evaluated on datasets encompassing up to 800 hip implant simulations. The finding showed that the trained models exhibit inadequate generalization capabilities, indicating that more data is necessary for the investigated problem formulation. To improve the performance for the given task, alternative encoding strategies or learning approaches may be necessary.     «

Design optimization poses significant challenges due to the substantial expensive and time-consuming characteristic of simulations. To counteract this issue, deep learning-based surrogate models have recently emerged as an effective solution. However, current research has primarily focused on applying these models to Computational Fluid Dynamics, with limited studies in the area of Linear Elasticity. Unlike previous work, the investigated structures are 3D with notable variations in the si...     »