I. Chamatidis, M. Stoumpos, G. Kazakis, N.A. Kallioras, S. Triantafyllou, V. Plevris and N.D. Lagaros, "Overview on Machine Learning Assisted Topology Optimization Methodologies", in Machine Learning in Modeling and Simulation (Part of the book series: Computational Methods in Engineering & the Sciences (CMES)), T. Rabczuk and K-J. Bathe (Eds.), Springer, pp. 373-394, 2023.

Abstract:
The past two decades saw tremendous developments in artificial intelligence (AI). Advancements in software, algorithms, and hardware led to the development of significantly more accurate and versatile artificial intelligence models. This rendered artificial intelligence a powerful tool that is used in diverse scientific areas, e.g. medicine and drug design, economics, and self-driving cars, among many others. These methods, having been successfully implemented in the simulation and modeling of structures, found their way to topology optimization problems, where artificial intelligence appears to have great potential for successful implementation. In conventional topology optimization, the optimal design of a specific domain must be calculated subject to specific constraints and the objective is to minimize the total compliance of the structure and use a specific amount of material. This is typically an iterative process that involves large matrices and can be very timeconsuming. By means of artificial intelligence models, referred to also as surrogate models (or surrogates), the computing time can be reduced significantly. The surrogate model is apriori trained offline. Following, during the optimization process the model is inferred based on input data, which is a lot faster due to limited matrix multiplications that the surrogate performs. The usual process involves either an artificial intelligence surrogate that complements the conventional procedure to reduce computational costs or a standalone surrogate which calculates the whole optimized structures by itself. The AI surrogates that are used belong to two main categories, i.e. Surrogates that use density and surrogates that use images. The surrogates that use density have similar inputs as the conventional method since the optimization process uses the density of the structure and is updated in each iteration of the AImodel. The surrogates that perform optimization on images are a bit different because they use techniques like image segmentation and filtering to output the optimized image (structure) which then is mapped into density. Most surrogates can be used for 2D and 3D structures and they are transferable, meaning that once trained they can be used in another topology optimization problem (thermodynamics or different material). The Background section contains an introduction to artificial intelligence, the surrogate models that will be used and an introduction to conventional topology optimization. The Literature Survey section provides areview ofrecent advancements of topology optimization using artificial intelligence models. This section is divided into two parts, the first describing the models that use density and the second the models that use image-based approaches.