Lauri Salmela

Great work has been released by Lauri Salmela @salmelala from the Ultrafast Photonics group @GGoery @flagshipprein 💡 The work has been also selected as an Editor's Pick by @OpticaWorldwide highlighting papers with excellent quality 💥 🔍 Check it out: opg.optica.org/ol/fulltext.cf…

@johnmdudley I have done some comparison between Python and Matlab only, and got worse results with Python at least in my case. But GNLSE solver in Fortran could be neat indeed

@salmelala Did you know about this?

@johnmdudley @GGoery @dbrunner_lab I got 10/10 :)

@GGoery @dbrunner_lab @salmelala 8/10 and I think the other 2 are wrong!

Here is something for @salmelala and @dbrunner_lab Will you score 100%?

@the_atman @GGoery @johnmdudley Hi Alexander. I have not really considered the interpretability of RNNs, and personally I don't see the opacity of RNNs as a problem at least with our research. Related to the preprint, I like the idea of hybrid models. Maybe this is something we could look into in the future

@GGoery @johnmdudley . @salmelala Very interesting work! Not a trick question, but what is your feeling regarding the physical interpretability of the various 'parameters' in the RNNs? And does it matter? Unrelated, but see: arxiv.org/abs/1810.02880

Check our latest research with @johnmdudley presented at POM on using a recurrent neural network to predict ultrafast nonlinear dynamics, from order soliton compression to octave-spanning supercontinuum generation. And using only the input pulse intensity as initial parameter.

@omairg @PhotonicsMeetup Hi. The fiber length could be changed freely here. Only issue may come from spectral or temporal broadering higher than in the samples used in the training if the length is increased too much. We only tested sech-typed input pulses but this approach it's limited to them only

We show how a recurrent neural network can predict the propagation dynamics of ultrashort pulses in an optical fiber for supercontinuum generation @PhotonicsMeetup #POM20ju #POM20AI

@Ariel_Levenson @PhotonicsMeetup Hi Ariel. For a single realisation, the gain in time depends on the simulation parameters (number of grid points) but is in our case around factor of 3-5 times faster. However, if you have multiple input conditions, the speedup is in the range of 20-50 times faster.

@salmelala @PhotonicsMeetup Very interesting ! Could you please quantify the gain in time for the full procedure?

@FlaviaTimpu @PhotonicsMeetup Hi Flavia, thanks for the question. We have here tested only one fiber (single-mode silica PCF), so the dispersion and the nonlinearity are always the same. In the training and testing samples, we have varied the input pulse peak power between 500W and 2kW with 100fs duration.

@salmelala @PhotonicsMeetup Great results! Can you tell me more about the training and testing samples, in particular if you tested for different properties of the fiber that may affect the SC (material, geometry)?

@sylvaingigan @PhotonicsMeetup Hi Sylvain. So far we have only tested networks with LSTM cells. Reservoir computing or other backpropagation algorithms (e.g. GRU cells) could be used as well. The choice between these options depends on the user's expertise.

@salmelala @PhotonicsMeetup very nice work ! did you try just with LSTM or did you compare with a simple reservoir computing paradigm? In short, is there a specific adavantage of the LSTM architecture ?