Search

A generative method inspired by leaf venation is presented for the layout design of heat conduction channels. This method, which is based on the simulation of the growth process of leaf venation, also employs the element-free Galerkin (EFG) method to discretize the design domain for the heat conduction structures. Using the EFG method, the growth of each cooling branch follows any arbitrary direction. In addition, a divergence criterion for the cooling branches is presented. The potential effectiveness and efficiency of this generative optimization design approach is investigated through a numerical examination of the “volume-to-point” heat conduction problem.

A deep learning approach combining with the traditional solid isotropic material with penalization (SIMP) method is presented in this paper to accelerate the topology optimization of the conductive heat transfer. This deep learning predictor is structured based on the deep fully convolutional neural network. The validity and accuracy of this deep learning approach is investigated based on the typical ‘Volume-Point’ heat conduction problems. The time consumption of the optimization process will be reduced significantly by introducing the deep learning approach.

A supervised deep learning predictor was developed to directly, intelligently, and instantaneously infer conductive heat transfer topologies. The architecture of the proposed supervised predictor consists of an encoder to decrease the dimensionality of the input data and a decoder. The high accuracy of the predictor is enabled by three parallel linked supervised deep learning predictors. Once the predictor has been trained and the physical parameters of the heat conduction problem, such as the boundary and constraint conditions, have been provided as input, the predictor directly and instantly outputs the optimized topology.