Machine Learning Accelerates Discovery of Optimized Fiber Optic Designs
A novel two-stage machine learning framework is proposed for the efficient identification of high-performance Nested Antiresonance Nodeless Fiber Designs (NANFs). This approach utilizes a neural network classifier for single-mode filtering and a regressor trained on the log of confinement loss (CL) to predict performance. The framework demonstrates the ability to explore vast design spaces ($14e6$ cases) computationally efficiently and has successfully identified designs with a CL of 0.25 dB/km from an initial set where CL $\ge$ 1 dB/km, indicating effective extrapolation and capture of underlying physical principles using limited data.