Scalable orthogonal delay-division multiplexed OEO artificial neural network trained for TI-ADC equalization

Abstract

We propose a new signaling scheme for on-chip optical-electrical-optical artificial neural networks that utilizes orthogonal delay-division multiplexing and pilot-tone-based self-homodyne detection. This scheme offers a more efficient scaling of the optical power budget with increasing network complexity. Our simulations, based on 220 nm silicon-on-insulator silicon photonics technology, suggest that the network can support 31×31 neurons, with 961 links and freely programmable weights, using a single 500 mW optical comb and a signal-to-noise ratio of 21.3 dB per neuron. Moreover, it features a low sensitivity to temperature fluctuations, ensuring that it can be operated outside of a laboratory environment. We demonstrate the network’s effectiveness in nonlinear equalization tasks by training it to equalize a time-interleaved analog-to-digital converter (ADC) architecture, achieving an effective number of bits over 4 over the entire 75 GHz ADC bandwidth. We anticipate that this network architecture will enable broadband and low latency nonlinear signal processing in practical settings such as ultra-broadband data converters and real-time control systems.

Wideband SiN pulse interleaver for optically-enabled analog-to-digital conversion: a device-to-system analysis with cyclic equalization

“…1(a) has five inputs, it can be implemented as a 5-by-5 multimode interferometer (MMI) without additional loss of power, so that the broadcasting over multiple Rxs can be implemented without further burdening the link budget. This is essentially the network topology already introduced in [16] for an on-chip ANN.…”

“…In the time domain, it can be inferred from the maximum of one pulse overlaying with zero-crossings of the others. As previously pointed out [16], this multiplexing scheme has some important differentiations relative to direct detection orthogonal time division multiplexing (OTDM). In particular, data can also be applied to fully dispersed and overlapping pulses, provided the subcarriers are delayed by the proper amounts in each of the branches.…”

“…We have previously shown that such an interferometer allows the generation of an almost arbitrary superposition of pulses, in this case spaced by multiples of 3𝜏 ! , with arbitrary pulse amplitudes and phases, provided the corresponding spectrum has comb line power levels below those of the comb lines at the input of the PPS [16].…”

“…Here, color-coding indicates pulses that have been modulated by the same phase-modulator, but that correspond to different sampling times. electrical (OEO) artificial neural networks (ANNs) and showed that it has a high robustness against environmental temperature drifts when implemented with a balanced network on a single chip [16]. ODDM transmission relies on frequency comb sources, but instead of mapping channels to individual comb lines as done in WDM systems [17]- [19], information is mapped over the entire comb spectrum while maintaining orthogonality between the channels.…”

“…Appendices I-III put the architecture explored here in a broader theoretical framework and analyze its robustness against group velocity dispersion (GVD). A comb source is first split between a reference branch carrying half the optical power and a second branch, that is additionally split between four independent modulators, as this allocation optimizes the optical power budget [16]. In addition, group delays are applied in each branch with a magnitude 𝜏 !…”