Summary
ALPI aims at the integration of a photonic neural network within an optical transceiver to increase the transmission capacity
of the optical link. Based on a deep learning approach, the new compact device provides real time compensation of fiber
nonlinearities which degrade optical signals. In fact, the tremendous growth of transmission bandwidth both in optical
networks as well as in data centers is baffled by the optical fiber nonlinear Shannon capacity limit. Nowadays, computational
intensive approaches based on power hungry software are commonly used to mitigate fiber nonlinearities. Here, we propose
to integrate in the optical link the neuromorphic photonic circuits which we are currently developing in the ERC-AdG
BACKUP project. Specifically, the proposed error-correction circuit implements a small all-optical complex-valued neural
network which is able to recover distortion on the optical transmitted data caused by the Kerr nonlinearities in multiwavelength
optical fibers. Network training is realized by means of efficient gradient-free methods using a properly designed
data-preamble.
A new neuromorphic transceiver demonstrator realized in active hybrid Si/InP technology will be designed, developed and
tested on a 100 Gbps 80 km long optical link with multiple-levels symbols. The integrated neural network will mitigate the
nonlinearities either by precompensation/autoencoding at the transmitter TX side or by data correction at the receiver RX
side or by concurrently acting on both the TX and RX sides. This achievement will bear to the second ALPI’s goal: moving
from the demonstrator to the industrialization of the improved transceiver. For this purposes, patents will be filed and a business plan will be developed in partnership with semiconductor, telecom and IT companies where a path to the commercialization will be
individuated. The foreseen market is the big volume market of optical interconnection in large data centers or metro
networks.
of the optical link. Based on a deep learning approach, the new compact device provides real time compensation of fiber
nonlinearities which degrade optical signals. In fact, the tremendous growth of transmission bandwidth both in optical
networks as well as in data centers is baffled by the optical fiber nonlinear Shannon capacity limit. Nowadays, computational
intensive approaches based on power hungry software are commonly used to mitigate fiber nonlinearities. Here, we propose
to integrate in the optical link the neuromorphic photonic circuits which we are currently developing in the ERC-AdG
BACKUP project. Specifically, the proposed error-correction circuit implements a small all-optical complex-valued neural
network which is able to recover distortion on the optical transmitted data caused by the Kerr nonlinearities in multiwavelength
optical fibers. Network training is realized by means of efficient gradient-free methods using a properly designed
data-preamble.
A new neuromorphic transceiver demonstrator realized in active hybrid Si/InP technology will be designed, developed and
tested on a 100 Gbps 80 km long optical link with multiple-levels symbols. The integrated neural network will mitigate the
nonlinearities either by precompensation/autoencoding at the transmitter TX side or by data correction at the receiver RX
side or by concurrently acting on both the TX and RX sides. This achievement will bear to the second ALPI’s goal: moving
from the demonstrator to the industrialization of the improved transceiver. For this purposes, patents will be filed and a business plan will be developed in partnership with semiconductor, telecom and IT companies where a path to the commercialization will be
individuated. The foreseen market is the big volume market of optical interconnection in large data centers or metro
networks.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/963463 |
| Start date: | 01-11-2020 |
| End date: | 30-04-2022 |
| Total budget - Public funding: | - 150 000,00 Euro |
Cordis data
Original description
ALPI aims at the integration of a photonic neural network within an optical transceiver to increase the transmission capacityof the optical link. Based on a deep learning approach, the new compact device provides real time compensation of fiber
nonlinearities which degrade optical signals. In fact, the tremendous growth of transmission bandwidth both in optical
networks as well as in data centers is baffled by the optical fiber nonlinear Shannon capacity limit. Nowadays, computational
intensive approaches based on power hungry software are commonly used to mitigate fiber nonlinearities. Here, we propose
to integrate in the optical link the neuromorphic photonic circuits which we are currently developing in the ERC-AdG
BACKUP project. Specifically, the proposed error-correction circuit implements a small all-optical complex-valued neural
network which is able to recover distortion on the optical transmitted data caused by the Kerr nonlinearities in multiwavelength
optical fibers. Network training is realized by means of efficient gradient-free methods using a properly designed
data-preamble.
A new neuromorphic transceiver demonstrator realized in active hybrid Si/InP technology will be designed, developed and
tested on a 100 Gbps 80 km long optical link with multiple-levels symbols. The integrated neural network will mitigate the
nonlinearities either by precompensation/autoencoding at the transmitter TX side or by data correction at the receiver RX
side or by concurrently acting on both the TX and RX sides. This achievement will bear to the second ALPI’s goal: moving
from the demonstrator to the industrialization of the improved transceiver. For this purposes, patents will be filed and a business plan will be developed in partnership with semiconductor, telecom and IT companies where a path to the commercialization will be
individuated. The foreseen market is the big volume market of optical interconnection in large data centers or metro
networks.
Status
CLOSEDCall topic
ERC-2020-POCUpdate Date
27-04-2024
Images
No images available.
Geographical location(s)
