A function online learning and prediction method for accelerating structural topology optimization and its application to pneumatic soft actuator

In this work, a novel higher-dimensional function online learning and prediction (FoLap) method is proposed to accelerate gradient-based and gradient-free topology optimization methods using machine learning. Firstly, a one-hidden layer neural network is proposed to learn and predict the update of a higher-dimensional function online. Secondly, the FoLap is proposed for designing compliant mechanisms considering design-dependent loads and fixed loads. Specifically, a bio-inspired thin-walled structure, named air pressure actuated horizonal expansion and lengthwise contraction (HELC) cell, is designed and optimized using the FoLap. The implementation of FoLap can save up to 65.1 % total computational time in the complete design process of a HELC cell, when compared with the selected level set method and the moving isosurface threshold method. Thirdly, the designed HELC cell can be assembled to build a pneumatic bending and contracting system (PBaCs), exhibiting capabilities for contraction, bending, and S-motions. The HELC cell and the PBaCs are prototyped via 3D printing and tested for its three actuations. The correlation between the experimental and nonlinear finite element analysis (NFEA) results validate the proposed FoLap method.