Prototyping a temperature-responsive flexible colorimetric sensor with machine learning-based result interpretation

Abstract

This research focuses on developing a flexible, colorimetric biosensor for temperature detection by integrating thermochromic nanofibers with machine learning for automated image-based analysis and temperature prediction. The sensor uses 10,12-pentacosadiynoic acid (PCDA), a material known for its visible blue-to-red transition in response to heat, embedded in a coaxial electrospun core–shell fiber. The outer shell, composed of PCDA, polyurethane (PU), and polyethylene oxide (PEO), enables a fast and visible color response, while the PEO-based core provides mechanical stability and supports fiber formation. The primary goal was to examine how different PCDA concentrations, temperature levels, and thermal exposure durations affect the intensity, speed, and consistency of the colorimetric response. Sensors were fabricated with PCDA concentrations ranging from 5% to 20% and tested across six temperatures (55°C to 80°C) and six exposure times (1 to 120 seconds). Among all combinations, the 18% PCDA concentration at 30 seconds produced the most vivid and stable color shift, making it the optimal configuration for practical use. To enable automated analysis, a convolutional neural network (CNN) was trained on over 4,000 images to predict temperature from RGB color data. The best model achieved a test R2 of 0.884 and a mean absolute error below 1.3°C, confirming its ability to accurately interpret thermochromic behavior. A web-based interface was also developed, allowing users to upload images and receive instant temperature predictions, eliminating the need for manual interpretation or lab-based equipment. The results demonstrate a strong correlation between material composition, stimulus conditions, and sensor performance, and show that machine learning can significantly improve both the accuracy and usability of colorimetric sensors. Real-world testing confirmed the sensor's flexibility, responsiveness, and visual clarity, making it suitable for integration into wearable safety gear, industrial equipment, and smart textiles. It also shows promise for use in food processing, packaging, and environmental monitoring—especially in settings where quick, low-cost temperature assessment is needed without electronic devices. Looking forward, the modular design of this platform offers clear pathways for expansion. Future work may explore using the same PCDA-based system for pH or bacterial detection, leveraging its inherent responsiveness, and potentially incorporating antimicrobial agents for dual-function sensing and drug release. Overall, this study lays the foundation for intelligent, low-cost biosensors that support real-time, automated monitoring in healthcare, environmental safety, and smart textile applications.