[Paper Review] A Recurrent Ising Machine in a Photonic Integrated Circuit

Mihika Prabhu, Charles Roques‐Carmes+11 more|arXiv (Cornell University)|Sep 30, 2019

Neural Networks and Reservoir Computing4 citations

TL;DR

This paper presents the Integrated Nanophotonic Recurrent Ising Sampler (INPRIS), a photonic integrated circuit that leverages optical matrix multiplication and controlled noise to efficiently sample the Gibbs distribution of Ising problems. The system achieves high-probability convergence to the ground state in 4-spin graphs by exploiting physical noise as a computational resource, demonstrating a path toward GHz-speed, low-power photonic Ising machines with potential orders-of-magnitude speedup over classical solvers.

ABSTRACT

Conventional computing architectures have no known efficient algorithms for combinatorial optimization tasks, which are encountered in fundamental areas and real-world practical problems including logistics, social networks, and cryptography. Physical machines have recently been proposed and implemented as an alternative to conventional exact and heuristic solvers for the Ising problem, one such optimization task that requires finding the ground state spin configuration of an arbitrary Ising graph. However, these physical approaches usually suffer from decreased ground state convergence probability or universality for high edge-density graphs or arbitrary graph weights, respectively. We experimentally demonstrate a proof-of-principle integrated nanophotonic recurrent Ising sampler (INPRIS) capable of converging to the ground state of various 4-spin graphs with high probability. The INPRIS exploits experimental physical noise as a resource to speed up the ground state search. By injecting additional extrinsic noise during the algorithm iterations, the INPRIS explores larger regions of the phase space, thus allowing one to probe noise-dependent physical observables. Since the recurrent photonic transformation that our machine imparts is a fixed function of the graph problem, and could thus be implemented with optoelectronic architectures that enable GHz clock rates (such as passive or non-volatile photonic circuits that do not require reprogramming at each iteration), our work paves a way for orders-of-magnitude speedups in exploring the solution space of combinatorially hard problems.

Motivation & Objective

• To develop a scalable, photonic-based Ising machine that efficiently solves NP-hard combinatorial optimization problems.
• To overcome limitations of existing physical Ising machines, such as low ground state convergence probability and poor universality for high-edge-density or weighted graphs.
• To leverage intrinsic and extrinsic noise in photonic circuits as a resource for faster exploration of the solution space.
• To demonstrate a proof-of-concept integrated nanophotonic system capable of operating at high clock rates with fixed, non-reconfigurable photonic circuits.
• To enable future quasi-passive photonic ASICs that outperform electronic and active photonic Ising machines in speed and energy efficiency.

Proposed method

• The system implements a recurrent Ising sampler using a programmable nanophotonic processor (PNP) that performs optical matrix multiplication with a fixed transformation matrix C = 2√K, where K is the coupling matrix of the Ising problem.
• Input spin states are encoded as in-phase optical signals and passed through the PNP, producing an output with Gaussian-distributed noise (standard deviation φ) due to physical imperfections.
• The noisy output is passed through an analog nonlinear unit (e.g., a saturating amplifier) before being fed back into the input, creating a recurrent loop that drives the system toward low-energy states.
• The system's dynamics are governed by a stochastic process that converges to the Gibbs distribution of the Ising Hamiltonian, with the ground state exponentially favored at low temperatures.
• Extrinsic noise is injected during iterations to enhance phase space exploration and improve convergence to the global minimum.
• The photonic circuit is designed to be non-reconfigurable per iteration, enabling GHz clock rates via passive or non-volatile photonic architectures.

Experimental results

Research questions

• RQ1 Can a photonic integrated circuit with fixed optical transformations efficiently sample the Gibbs distribution of arbitrary Ising problems?
• RQ2 How does controlled physical noise enhance convergence to the ground state in a recurrent Ising sampler?
• RQ3 To what extent can intrinsic and extrinsic noise be leveraged as a computational resource in photonic Ising machines?
• RQ4 Can a non-reconfigurable, passive photonic architecture achieve high-speed, low-energy computation for combinatorial optimization?
• RQ5 What are the key design trade-offs in scaling up such a system to larger spin networks?

Key findings

• The INPRIS demonstrated high-probability convergence to the ground state in various 4-spin Ising graphs, confirming its ability to sample the Gibbs distribution accurately.
• The system achieved convergence to the ground state with high success probability by exploiting physical noise as a resource, with noise levels near the optimal regime for ground state search.
• Experiments showed that both intrinsic and extrinsic noise sources significantly influenced physical observables such as magnetization and ground state population, confirming noise-dependent dynamics.
• The recurrent photonic transformation is fixed per problem instance, enabling potential implementation in non-volatile photonic ASICs with GHz clock rates and attojoule-level energy consumption per step.
• The system's performance was validated through simulations and experiments, showing convergence to the Gibbs distribution as predicted by statistical mechanics.
• The work identifies key design trade-offs, such as eigenvalue dropout and idler signal count, that must be addressed in scaling up to larger systems.