Abstract
This paper proposes a high-sensitivity-integrated temperature sensor with low complexity based on a silicon waveguide. The waveguide layout is optimized through the finite-difference time-domain (FDTD) simulations, and a compressed taper structure improves the efficiency of speckle data collection while reducing the system complexity and cost. To achieve precise temperature demodulation, this paper employed a convolutional neural network (CNN) for nonlinear fitting. Experimental results demonstrate the sensor's ability to perform temperature measurement in the range of -20 °C to 100 °C, with a best resolution of 0.00287 °C (2.87 mK). The resolution and reliability of the measurements are validated by comparison with the theoretical values. This study introduces a novel approach to silicon waveguide-based temperature sensing.