Machine learning empowers large-scale optical sensors for ultrasensitive detection

Optical sensors are stirring broad interests in disease diagnostics, food safety, and environment monitoring [1, 2, 3]. Several criteria assess the performance of a sensor, including the analytical detection speed, cost, sensitivity, and reproducibility [4, 5]. Traditionally, optical sensing leverages localized spectral features such as e.g., resonance peaks shift, intensity variations, and widths. This approach, while straightforward in implementation, results in a weak detection limit for analytes, and needs improvement for enabling practical applications. Recent pioneering work focuses on artificial intelligence (AI) to address this issue, leveraging sparse features in broad amounts of data to enhance the sensor detection sensitivity [6]. However, most of these approaches rely on post-processing data collected with complex equipment, such as spectrum analyzers. These systems are significantly expensive, not integrated, and compete poorly with traditional sensing based on localized features [7].