Customizing Tactile Sensors via Machine Learning-Driven Inverse Design

Replicating the sophisticated sense of touch in artificial systems requires tactile sensors with precisely tailored properties. However, manually navigating the complex microstructure-property relationship results in inefficient and suboptimal designs. Here, we present a machine learning-accelerated, multi-objective inverse design methodology for the automatic customization of tactile sensors. At its core is a data-efficient microstructure-property predictor designed to ensure robust accuracy with minimal experimental data. It achieves this by synergistically combining support vector machine-based boundary definition with dual-phase active learning. This predictor then drives a multi-objective inverse design software, enabling real-time, on-demand sensor customization. This methodology not only dramatically enhances the design efficiency but also yields sensors with exceptional characteristics-high sensitivity (1.2 V/kPa), high linearity (R² = 0.999), and wide detection range (0-400 kPa). The resulting sensors are successfully applied to pulse monitoring, material identification, and robotic grasping. Furthermore, the underlying microstructure-property mechanisms are elucidated using interpretable machine learning. This work establishes a general paradigm for automated sensor customization, accelerating the development of next-generation wearable and robotic sensing systems.