Photonics Research, Volume. 12, Issue 1, 85(2024)
Scalable orthogonal delay-division multiplexed OEO artificial neural network trained for TI-ADC equalization
Fig. 1. Orthogonal delay-division multiplexing scheme represented for
Fig. 2. Signal weighting schemes. In (a), phases can be set independently in the top demodulator branch at the expense of reducing the amplitude proportionally to the number of paths. In (b), this is alleviated by introducing arrayed waveguide gratings to (de-)multiplex the comb lines, allowing, in principle, lossless signal recombination. However, AWGs require a substantial chip area. Significant reduction of the demodulator size is achieved by simultaneously allowing the pilot tone to cross each path in the upper demodulator branch of (c) so that a programmable superposition of all possible combinations of delays is obtained. (d) Further simplified network in which the embedded MZIs have been replaced by static splitters with an off-3-dB splitting ratio.
Fig. 3. Numerical trials for the validation of the weight space addressable by the demodulator architecture shown in Fig.
Fig. 4. Diagram of a
Fig. 5. Exemplary simulation result with a comb FWHM spanning 75% of
Fig. 6. Signal distortion (std. dev. of the difference between the demodulated photocurrent and the transmitted signal) versus (a) comb FWHM and (b) Bessel filter cutoff frequency. In (a), the Bessel filter cutoff is maintained at 6.25 GHz, and in (b), the FWHM is maintained at
Fig. 7. Simulated noise std. dev.
Fig. 8. (a) Simulated noise std. dev.
Fig. 9. (a) Schematic of the auxiliary SiN chip generating the 25 GHz comb shared between the OE-TI-ADC front-end and the OEO-ANN. Prior to routing it to the ANN, the FSR of the comb is doubled with an imbalanced MZI. (b) Six-channel OE-TI-ADC front-end. Interleaved pulse trains with different center frequencies are fed to a dual-output MZM, after which the pulse trains are de-interleaved by a wavelength division demultiplexer. (c), (d) Schematics of the ANN input layer. In (c), required delays are implemented in the electrical domain, prior to electro-optic phase modulation. In (d), signals are split and delayed in the optical domain instead, with delays chosen such that the delayed optical signals are mapped to free ODDM channels. (e) Higher-level schematic of the overall ANN, with two optical networks connecting, respectively, the input to the hidden layer (
Fig. 10. (a) Photocurrent at the output of the analog output ANN as a function of its input (bottom axis). The nonlinearity of the system is apparent. The ANN corrects for the OE-TI-ADC front-end nonlinearity, as shown by the red curve. It corresponds to the output of the ANN as a function of the input to the OE-TI-ADC (top axis) and has a very linear profile. In (b), the photocurrent of every third output neuron of the digital output ANN is shown as a function of the input to the ANN. The crossing of these curves with the zero differential photocurrent level corresponds to the switching threshold of the following 1-bit digitizer.
Fig. 11. (a) SNR levels of the processed signals in different scenarios, highlighting the improvement with respect to the non-equalized signal, also in presence of noise. The SNR is assessed by processing sinusoidal signals at the frequencies reported on the abscissa. In (b), the mean SNR, assessed by averaging the variances of signals at multiple frequencies, is shown as a function of different rescaling factors applied to the std. dev. of the noise terms. The analog output ANN is seen to be significantly more sensitive to noise than the digital output ANN.
Fig. 12. DCS design and characteristics. (a) Layout of the DCS designed into the 220 nm SOI device layer, (b) power transmission coefficients, and (c) group delays associated with the four port combinations labeled in (a) after balancing with external waveguide segments sized for 1550 nm.
Fig. 13. Star coupler design and characteristics. (a) Layout of the star coupler with overlaid field intensity for light injected through the port of index
Fig. 14. Floorplan of a 31-by-31 ANN-PIC. The insets show details of the layout.
Get Citation
Copy Citation Text
Andrea Zazzi, Arka Dipta Das, Lukas Hüssen, Renato Negra, Jeremy Witzens, "Scalable orthogonal delay-division multiplexed OEO artificial neural network trained for TI-ADC equalization," Photonics Res. 12, 85 (2024)
Category: Integrated Optics
Received: Apr. 24, 2023
Accepted: Oct. 9, 2023
Published Online: Dec. 21, 2023
The Author Email: Jeremy Witzens (jwitzens@iph.rwth-aachen.de)
CSTR:32188.14.PRJ.493888