Predicting the adhesion strength of anisotropic microstructures using Physics-Informed Tandem Neural Networks

In this study, we propose a Physics-Informed Tandem Neural Network (PI-TNN) framework for accurate adhesion prediction and robust geometric optimization. To address the non-uniqueness of inverse mapping, a tandem architecture is constructed by coupling an inverse generator with a pre-trained forward predictor. In addition, a power-law hardening constraint ( n ≈ 2.04 ) derived from contact mechanics is incorporated into the loss function as a physical regularization term, improving physical consistency under limited training data. An Adhesion Structure Factor ( S ) is further introduced to quantify the geometric-elastic contribution independently of interfacial chemical energy. The proposed framework is validated using high-fidelity finite element simulations and experiments on micro-wedge arrays fabricated by pulsed laser ablation. The results show good agreement with both experimental observations and literature data, and indicate that adhesion performance is governed more strongly by structural geometry than by material elasticity. This work provides an effective surrogate model for nonlinear elastic microstructures and a practical framework for the design of high-performance soft adhesives.