Machine Learning–Driven Inverse Design of Gradient Porous Elastomers for High-Performance Flexible Pressure Sensors

Abstract Introducing gradient porous microstructures offers a cost-effective and easily scalable approach for developing flexible thin-film pressure sensors. However, the vast design space arising from varying pore sizes and combinations renders the identification of optimal structural parameter nearly unattainable through conventional trial-and-error experiments. Herein, we propose a novel methodology that integrates finite element simulation with machine learning to efficiently search for the optimal solution. A finite element model of the gradient porous dielectric layer is established using COMSOL, incorporating an equivalent parameter inversion strategy. Hundreds of pressures–capacitance response datasets are generated through parametric simulations and subsequently employed to train a backpropagation (BP) neural network surrogate model. The optimized structural parameters are then experimentally validated through the fabrication of a thin-film pressure sensor. The fabricated sensor exhibits fast response and recovery time (75 ms), a low detection limit (0.2 kPa), a wide detection range (600 kPa), and outstanding long-term stability over 5,000 cycles. Furthermore, the sensor demonstrates excellent applicability in various scenarios, including human physiological signal monitoring and speech recognition. More importantly, the integrated simulation–machine learning framework presented herein is broadly applicable to other structural designs, offering a powerful alternative to traditional trial-and-error approaches and substantially reducing development costs.