Supercontinuum light is generated by a train of laser pulses propagating in an optical fiber. The parameters characterizing these pulses influence the spectrum of the light as it exits the fiber. While spectrum generation is a direct process governed by nonlinear equations that can be reproduced through numerical simulation, determining the parameters of the pulse generating a given spectrum is a difficult inverse problem. Solving this inverse problem has a relevant practical implication, as it allows generating beams with desired spectral properties. We solve this multidimensional parameter estimation problem by training a neural network and we introduce, as key technical contribution, a weighted loss function that improves the estimation accuracy. Most remarkably, this loss function is not specific to the considered supercontinuum scenario, but has the potential to improve solutions of similar inverse problems where the forward process can be reproduced via computationally demanding simulations. Our experiments demonstrate the effectiveness of the pursued approach and of our weighted loss function.

A Weighted Loss Function to Predict Control Parameters for Supercontinuum Generation Via Neural Networks

Stucchi Diego;Corsini A.;Boracchi G.;
2021

Abstract

Supercontinuum light is generated by a train of laser pulses propagating in an optical fiber. The parameters characterizing these pulses influence the spectrum of the light as it exits the fiber. While spectrum generation is a direct process governed by nonlinear equations that can be reproduced through numerical simulation, determining the parameters of the pulse generating a given spectrum is a difficult inverse problem. Solving this inverse problem has a relevant practical implication, as it allows generating beams with desired spectral properties. We solve this multidimensional parameter estimation problem by training a neural network and we introduce, as key technical contribution, a weighted loss function that improves the estimation accuracy. Most remarkably, this loss function is not specific to the considered supercontinuum scenario, but has the potential to improve solutions of similar inverse problems where the forward process can be reproduced via computationally demanding simulations. Our experiments demonstrate the effectiveness of the pursued approach and of our weighted loss function.
IEEE International Workshop on Machine Learning for Signal Processing, MLSP
978-1-7281-6338-3
File in questo prodotto:
File Dimensione Formato  
2021_02_MLSP_Stucchi.pdf

Accesso riservato

: Publisher’s version
Dimensione 734.26 kB
Formato Adobe PDF
734.26 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: http://hdl.handle.net/11311/1203791
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact