Sketch driven machine-learning based topology optimization

Abstract

Sketch design plays a very important role in model design. In order to improve the efficiency of existing design models that rely on computer-aided and human experience guidance, this work proposes a sketch driven machine-learning based topology optimization method. It helps designers directly design hand-drawn sketches to obtain topology-optimized structures that conform to sketching experience. The proposed method uses neural style transfer technique, and can compensate for the lack of design experience to obtain optimized structures without the need of multiple computational simulation interactions. Specific structural shapes and design styles according to the design requirements also can be obtained. In contrast to the approach of specifying the undesignable domain and initial layout, similarity constraints between sketches and structures are constructed to quantify the degree of inheritance of different sketches. Both 2D and 3D problems are solved to illustrate the effectiveness of the proposed approach.