Motivation
The beam of a laser is largely defined by the transverse modal profiles of the lasing modes. In cases where the structure and theory of a laser is well understood, one can obtain good estimates of the profiles from theory or waveguide/cavity mode-solving simulation. However, in some novel designs one may not be able to obtain reasonable estimates from theory or simulation. In these cases we can obtain modal profiles from experiment.
If we have a single (transverse) mode laser, then simply imaging the near-field of the laser beam should be sufficient. However, if we have a multi-mode laser then the problem becomes more difficult as the near-field will be a summation of the various mode profiles. One can separate the modes and image them using varied focus or diffraction, but this adds experimental and equipment complexity.
Now, what if we could experimentally estimate the modal profiles from a set of multi-mode images?
The Method
Fig 1: Analyzing simulated mode recovery using an autoencoder
Now, we assume that we have a “training” dataset of multi-moded near-fields for which we know the modal powers (that is, we know what modes are present and how much power is in each mode). We can use an autoencoder artificial neural network, a combination of encoder and decoder networks, to encode the images into the modal power vector, and them decode them back to the images. We train the encoder network to minimize the modal decomposition losses (the difference the encoder’s prediction of the modal powers and the measured powers), and we train the decoder to minimize the reconstruction losses (difference between the decoder’s predicted near-field image and the measured image).
Once the autoencoder is trained, we have two separate networks. The encoder is capable of modal decomposition, so we can use it to estimate how much of each mode is present in a multi-mode near-field image. The decoder is capable of modal recovery, so if we ask it for a near-field images would be for situations where all of the power is in a single we can estimate the modal profiles of the individual modes.
Resources
- “Machine Learning for Modal Analysis” (DOI) (IEEEXplore) (Abstract) (Presentation Slides)
- ModeAnalysisML.jl: Julia package that implements machine learning based modal decomposition and recovery