Fig. 2: Improving Emission Directivity with Active Learning.
a Top (bottom) panel shows the active learning result for the emission angle at +(−) 26.6°. Blue dots represent training points (Sobol sampling) in the VAE latent space, and black points represent improved steering signals learned by the active learning agent. The blue dashed line indicates the average of the training points, while the black line indicates the average of the learning points. Red vertical stripes show measurement errors averaged over 10 repeats. b Relative improvement in peak directivity by the active learning agent across multiple far-field emission angles through closed-loop experimental feedback. The top panel shows the peak directivity of emission optimized for each far-field angle, with the active learning agent’s values in blue and the saw-tooth grating values in red. The bottom panel shows the relative improvement in peak directivity for different emission angles enabled by the active learning agent. c Left panel shows the optimal VAE output (normalized) learned by the active learning agent, and the right panel shows the normalized 1–D pump pattern transforming the VAE output for loading into the SLM: YSLM = (YVAE % 2π)/2π for different emission angles.
