Real-Time Machine Learning-Based Position Recognition in Laser Nanofabrication with Sub-Half-Wavelength Precision

Laser nanofabrication with tightly focused ultrafast laser pulses enable versatile fabrication of arbitrary two-dimensional (2D)/three-dimensional (3D) micro/nanostructures. Accurate positioning of the laser focal spot is crucial, especially for high-resolution integrated devices in 2D materials, requiring precise placement on atomically thin surfaces. However, uneven surfaces and surface tilt poses significant challenges. Existing methods to detect focal positions often involve complex setups with additional optical components or sensors, achieving limited accuracy. This study introduces a machine learning-based method to accurately detect the focal position during laser nanofabrication by analyzing the shape and intensity of the laser focal spot. We compare four machine learning methods: rational quadratic Gaussian process regression, kernel approximation least square, quadratic support vector machine, and trilayered neural network. Our experiments show that TNN method achieves a detection accuracy of 257 nanometers (nm) (half the fabrication laser wavelength), surpassing the required accuracy for tightly focused laser beam (typically larger than one wavelength). This method can map focal positions along the fabrication trajectory to compensate for surface roughness or tilt. Moreover, it can be directly implemented in any laser nanofabrication system with a camera for in situ monitoring, without requiring additional optical components, indicating broad applicability.