Machine learning-based optimization of the design of composite pillars for dry adhesives

Abstract

A machine learning-based optimization strategy was used to find optimal designs for adhesive composite pillars with either high adhesion strength or high adhesion tunability. Neural networks were trained with data generated by finite element analysis to predict the adhesion strength of composite pillars with different designs; an average prediction error of less than 1% was achieved. Through a sensitivity study with the trained neural networks, it is found that the geometry of the stiff core above a critical cut off height has no effect on the interfacial stress distribution and the adhesion of the pillar. A randomly initialized constrained optimization algorithm was then applied to the trained neural networks to find the optimal composite pillar design. A composite pillar design with a stiff core that has an enlarged tapered flat end is optimal for realizing robust and high adhesion, since it can achieve high adhesion under different loading and contact conditions. The optimized pillar has a critical normal detachment force that is nearly 11 times that of a homogenous pillar and 1.7 times that of a composite pillar with a simple wide rectangular core under normal loading. A composite pillar with a thin flat stiff core shows the highest effective adhesion difference between being loaded with a normal force and being loaded with a moment.