Voxel meshes are a natural choice for several types of analyses, particularly in the context of certain (gradient-based) optimization methods (including machine learning) or biomimicry. Such meshes can not directly be used to calculate meaningful surface stresses, though, e.g. for modeling fatigue. In this work, a process for calculating stresses based on neuronal nets (instead of smoothing and/or remeshing) is proposed, thereby potentially enabling the integration of stress calculation into optimization methods that rely on backpropagation (including generative design approaches). The method is demonstrated for 2D plane stress problems. Verification is attempted and the concrete behavior as well as limitations of the current implementation are discussed. © 2021 CAD Solutions, LLC.
