Abstract:A method is created to automatically increase the threshold projection parameter in three-field density-based topology optimization to achieve a near binary design. The parameter increase each iteration is based on an exponential growth function, where the growth rate is dynamically changed during optimization by linking it to the change in objective function. This results in a method that does not need to be tuned for specific problems, or optimizers, and the same set of hyper-parameters can be used for a wide range of problems. The effectiveness of the method is demonstrated on several 2D benchmark problems, including linear buckling and geometrically nonlinear problems.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to achieve a near - binary design in topological optimization. Specifically, the author proposes a method to achieve the near - binary design goal by automatically increasing the threshold projection parameter in the three - field - density - based topological optimization process. The core of this method is to dynamically adjust the growth rate of the threshold projection parameter according to the rate of change of the objective function at each iteration, so that the optimization process does not need to adjust parameters for specific problems or optimizers, and the same set of hyper - parameters can be applied to a wide range of problems.
### Background and Motivation of the Paper
Topological optimization aims to find the ideal layout, shape, and size of a structure under given objectives and constraints. Such problems are usually expressed by determining which positions in the design domain should have materials, thus forming a binary problem, that is, the material exists (with a value of 1) or does not exist (with a value of 0). However, solving binary optimization problems is challenging, so continuous design variables are usually used to relax the problem, such as the SIMP (Simple Isotropic Material with Penalization) method, in which each element is assigned a pseudo - density value between 0 (empty) and 1 (solid).
Although the pseudo - density method combined with filtering techniques can avoid numerical problems, such as checkerboard patterns and mesh - dependence, this method may lead to solutions containing a large number of intermediate pseudo - density values (between 0 and 1), far from the ideal binary design. To force the design to be close to binary, a threshold projection technique can be used, setting the part with a pseudo - density value higher than the threshold to 1 (solid) and the part lower than the threshold to 0 (empty). However, this introduces an optimization hyper - parameter that needs to be selected by the user - the projection parameter, which defines the sharpness of the projection. Empirically, many researchers adopt a continuation scheme for the projection parameter, that is, keeping a small value in the early iterations to allow the design to gradually form, and then gradually increasing it to push the design towards a binary solution.
### Proposed Method
This paper proposes a new method to automatically adjust the projection parameter by linking the increase of the threshold projection parameter to the progress of the objective function. Specifically, at each iteration, the increase amount of the projection parameter is determined according to the relative rate of change of the objective function. If the objective function is still changing significantly, it indicates that the design is still in the formation stage, and at this time, the speed of increasing the projection parameter is slow; if the change of the objective function tends to be stable, it indicates that the design has been basically formed, and at this time, the projection parameter can be increased more quickly to push the design towards a binary solution.
### Experimental Verification
The author verifies the effectiveness of this method through several 2D benchmark problems, including linear buckling and geometric nonlinear problems. The experimental results show that, compared with the traditional method of manually adjusting the projection parameter, the proposed method can reduce the gray area (i.e., the intermediate density value), and in some cases, can also reduce the number of required iterations.
### Conclusion
The method of automatically adjusting the threshold projection parameter proposed in this paper can effectively achieve a near - binary topological optimization design without sacrificing design performance. Compared with traditional methods, this method reduces the need for manual parameter adjustment for specific problems and has better universality and robustness. Future work will further test the application of this method in a wider range of problem types, including different physical phenomena and the use of different optimizers.