Machine learning-enabled performance prediction and design for flexible strain sensors

Developed an interpretable machine learning framework for carbon-based flexible strain sensors. • Multilayer Perceptron (MLP) achieved excellent prediction accuracy (maximum R 2  = 0.9600). • Comparative evaluation of MLP, RF and SVR modeling approaches for flexible strain sensor performance prediction. • SHAP analysis identified the core process parameters and synergistic mechanisms regulating sensor performance. • This framework supports the data-driven design and optimization of high-performance carbon-based flexible strain sensors. The sensing performance of carbon-based flexible strain sensors is largely governed by processing parameters. Therefore, the exploration of optimal processing parameters is of great significance to the enhancement of material properties. However, the complex coupling among multiple variables renders conventional trial-and-error optimization inefficient. Here, a data-driven prediction–optimization strategy integrating design of experiments (DoE) with machine learning (ML) is established to accelerate the regulation of sensing performance in a thermoplastic polyurethane/carbon nanotube-carbon fiber/thermoplastic polyurethane (TPU/CNTs-CF/TPU) strain sensor. Uniform and information-rich experimental datasets are constructed within a high-dimensional parameter space, and a multilayer perceptron (MLP) model is trained to characterize the relationship between processing parameters and sensing performance. SHAP-based interpretability analysis is further employed to quantitatively elucidate the contribution mechanisms of key factors, including carbon loading and carbon nanotube ratio, thereby guiding targeted experiments in high-sensitivity regions and enabling iterative optimization through dataset expansion. As a result, the sensor sensitivity is markedly improved, with the gauge factor (GF) increasing from 15 to 105 while maintaining high predictive accuracy (R 2  = 0.96). The trained model is subsequently used to screen 600 randomly sampled parameter combinations, from which two optimal candidates are identified and experimentally validated, yielding relative prediction errors as low as 2.3% and 3.1% at 80% strain. Moreover, the optimized sensors exhibit excellent stability under cyclic loading and human-motion monitoring. This work provides an efficient and interpretable paradigm for regulating material–process–performance relationships in flexible electronics, substantially reducing experimental cost and offering broad methodological guidance for high-performance flexible sensing systems. Download: Download high-res image (373KB) Download: Download full-size image