LightPro: A Linear Photonic Processor with Full Programmability

TL;DR

LightPro introduces a fully programmable linear photonic processor leveraging phase-change material–tunable directional couplers to enable scalable, low-footprint, and low-power MVM for photonic AI accelerators. A neural architecture search (NAS) with a fidelity-based pruning/snapping workflow optimizes the topology to realize arbitrary unitary transformations, achieving large reductions in footprint and active power while preserving accuracy on benchmark datasets. The work provides detailed device design (PCM DCs, TiN heaters), coupled-mode theory modeling, and Lorentz-model-based material simulations, complemented by experimental validation on an iPronics SmartLight platform with strong agreement to simulations. Overall, LightPro demonstrates a viable path toward scalable, energy-efficient photonic MVM hardware, with potential enhancements from all-PCM implementations and further reductions in crosstalk and calibration overhead.

Abstract

In this paper, we propose a novel fully programmable linear photonic processor, which we call LightPro, with improved scalability, performance, and footprint. At the heart of LightPro are compact, low-loss, and programmable silicon photonic (SiPh) directional coupler (DC) devices that deploy phase-change material (PCM) for programming the DC's splitting ratio. By thermally inducing phase transitions in the PCM, the coupling coefficient of the DC can be dynamically adjusted to achieve different splitting ratios in the device output. Building on this device foundation, we develop a neural architecture search (NAS) and pruning algorithm to optimize the architecture of the processor for performing MVM operations. Our simulation results show that LightPro achieves up to an 85% reduction in footprint and more than 50% improvement in power consumption. In addition, LightPro is evaluated by performing inference with weight matrices trained on MNIST and linearly separable Gaussian datasets, showing less than a 5% drop in accuracy when scaling up the network. Prototyping results, using a commercial photonic processor (iPronics SmartLight), show LightPro's efficiency and performance (e.g., computational accuracy) compared to conventional photonic MVM hardware, demonstrating the experimental evaluation and feasibility of LightPro for next-generation photonic AI accelerators.

Figures (9)

• Figure 1: Working principle of LightPro. The complex weights of a DNN trained on a fast Fourier transformed dataset can be decomposed using singular value decomposition amin_jltbanerjee2021champ. The decomposed weight matrix can be implemented as the multiplication of two complex unitary matrices and one diagonal matrix. The complex unitary matrices can be used as a target transfer matrix in LightPro. Using the target unitary matrix, LightPro employs a NAS to optimize the network topology, leveraging columns of PCM-based tunable DCs and phase shifters to carry out MVM in the optical domain. The splitting ratio of the DCs used in the linear multiplier network can be adjusted by changing the phase state of the PCM.
• Figure 2: (a) Schematic of the tunable PCM-based DC as well as its design parameters (see the right hand side figure). A TiN microheater is used to heat the PCM and induce phase change. Exploration of effective refractive index for a Sb$_2$Se$_3$-loaded silicon waveguide with different offset and width values when the Sb$_2$Se$_3$ is in the (b) amorphous state and (c) crystalline state. (d) CMT simulation results of the coupling coefficient for an example design point where W$_{PCM}$ = offset + W$_2$, W$_1$ = 0.450, $\mu$m, W$_2$ = 0.408 $\mu$m and offset = 100 nm when the PCM is in amorphous and crystalline state. (e) The corresponding EME verification simulation for the design picked in (d) for Sb$_2$Se$_3$ has a different phase states ($X_f = 0$: Amorphous state, $X_f = 1$: crystalline state), (f) Temperature distribution of the designed heater along its length when a reset heat pulse is used to melt the Sb$_2$Se$_3$ and change its phase state to amorphous state shafiee2024programmable.
• Figure 3: An overview of the proposed progressive optimization based on NAS to find the best network configuration in LightPro. Here, $\kappa_{min, max}$ can be set to be 0 and 1 for the initial progressive optimization. (b) Proposed pruning to remove phase shifters from the general-purpose LightPro network, to perform MVM for a specific application (i.e., when the unitary matrix is fixed). (c) Proposed snapping method to replace DCs with waveguide crossing or straight waveguide. (d) Left: An example of a 4$\times$4 LightPro network optimized using NAS, and right: the equivalent pruned LightPro network. (e) Left: An example of a 8$\times$8 LightPro network optimized using NAS, and right: the equivalent pruned LightPro network. In (d) and (e), the green boxes show PCM and the red ones show thermo-optic phase shifters.
• Figure 4: Phase shift and coupling coefficient distribution of the NAS optimized networks after the progressive optimization and the pruning step. The left hand side figures show the distribution of phase shifts in the network at different stages of optimization and pruning, while the right-hand side figures show the coupling coefficients after progressive optimization of the network configuration and snapping stages.
• Figure 5: (a) The boxplots denoting the fidelity of the NAS optimized network of different sizes when its parameters (phase and coupling coefficients) are re-adjusted to implement 100 different random complex unitary matrix. The bold red number denotes the average achieved fidelity amongst 100 random and complex unitary matrices. (b) The fidelity of the NAS optimized network with different sizes after progressive optimization to get the optimized configuration of the network, after phase shifter pruning, after the DC snapping and final parameter re-optimization.(c) Footprint area in square millimetres for the NAS optimized network and its comparison to its equivalent MZI-based Clements network. (d) The total number of active components for the NAS optimized and pruned networks and their comparison to the MZI-based Clements network with different sizes. (e) Total active power consumption related to the phase shifters for NAS optimzied and pruned networks and their comparison with the MZI-based Clements network with different sizes when the same multiplication was implemented. Note that the phase shifters are assumed to be based on thermo-optic effect. (f) Accuracy results of pruned NAS network of different sizes when they are used to re-caclulate the accuracy of the network trained on a linearly separable Gaussian dataset. The last bar denotes the case where a 3 hidden layers with SVD configuration trained on FFT-MNIST when the original weights and the pruned weights are used to calculate the accuracy.