Photonics Research Vol. 13, Issue 7, (2025)

Programmable photonic integrated circuits (PICs) achieve precise control over optical transmission paths through the dynamic tuning of unit structures. This capability enables software-defined, intelligent routing of optical signals, thereby supporting real-time analog signal processing. Thanks to this unique advantage, programmable PICs demonstrate broad application prospects in fields such as wavelength routing, optical neural networks, and microwave photonics. Therefore, with increasingly complex and diverse application scenarios, the trend toward large-scale integration of programmable PICs has become inevitable. This brings a critical challenge: how to achieve global optimal configuration of hundreds to thousands of control units to meet the demands of multifunctional signal processing. Traditional optimization algorithms (such as the Dijkstra algorithm, genetic algorithms, etc.) often struggle with effective convergence due to their exponentially increasing computational complexity. Therefore, it is imperative to develop intelligent computing models that align with the structural characteristics of PICs, enabling significant enhancements in system reconfiguration capabilities and accelerating the practical deployment of large-scale programmable PICs.

In response to this challenge, Professor Li Ming and Academician Zhu Ninghua's research team at the Institute of Semiconductors, Chinese Academy of Sciences, has made a major breakthrough in the field of AI-driven intelligent programmable photonic chips. The team proposed a novel Ising-model-based intelligent computing scheme for the reconfiguration of photonic integrated circuits. This approach leverages the high-efficiency optimization capabilities of the Ising model to achieve millisecond-level dynamic reconfiguration of programmable PICs. The researchers mapped the transmission matrices of photonic unit devices to Ising spin states, thereby constructing a complete Ising model for programmable PICs. Experimental results demonstrated that this scheme successfully reduced the reconfiguration optimization time of a 56×56 large-scale programmable PIC (with over 2,000 unit devices) to 30 milliseconds. Functionality verification through all-optical routing switching and optoelectronic neural networks confirmed the practicality and reliability of the proposed method. The relevant research results are published in Photonics Research, Volume 13, Issue 7, 2025. [Menghan Yang, Tiejun Wang, Yuxin Liang, Ye Jin, Nuannuan Shi, Ming Li, "Scalable and rapid programmable photonic integrated circuits empowered by Ising-model intelligent computation," Photonics Res. 13, 7 (2025)]

Figure 1 illustrates the construction principle of the equivalent Ising model for programmable PICs. Figure 1(a) shows a conceptual diagram of a programmable PIC, where the core tuning unit adopts a Mach–Zehnder Interferometer (MZI) structure. By precisely adjusting the state parameters of each MZI, dynamic programming of the light propagation path can be achieved. Figure 1(b) presents an innovative Ising model equivalence method for the unit structure: each physical state of an MZI unit is mapped to a combination of four Ising spins, thereby establishing a one-to-one correspondence between the entire photonic integrated circuit and the Ising model. Based on this model, the Ising intelligent solving algorithm shown in Figure 1(c) is applied to obtain the optimal reconfiguration scheme for the programmable PICs.

Fig. 1. Programmable PICs and the equivalent Ising model. (a) The structure of programmable PICs consists of a cascading arrangement of hexagonal structures, with each MZI as a programmable basic unit. The MZI is designed with a phase shift in one arm, enabling precise control over the optical power. (b) The correspondence states between the MZI and the binary decision variable. The MZI's cross state at θ=0 corresponds to the node state '0', whereas the MZI's bar state at θ=π is equated with the node state '1'. (c) The equivalent Ising network model and Ising-model intelligent computation process. A complete Ising matrix is formulated by incorporating the constraints matrix into the Ising model and then finding the optimal solution via an Ising machine to enable dynamic control over the PICs. MZI: Mach–Zehnder interferometer, PS: phase shifter

Through this equivalence-based computational method, programmable PICs can be mapped to an Ising model, effectively transforming the reconfiguration and optimization problem of PICs into one of finding the global minimum energy state of the corresponding Ising system. When the energy converges to its lowest point, the optimal configuration for chip reconfiguration is obtained. This Ising-model-based intelligent computation, when integrated with emerging computing architectures such as the microwave photonic Ising machine (MPIM), dramatically reduces optimization time to the millisecond level, marking a qualitative leap in computational efficiency. Notably, in simulating a 56×56 large-scale PIC system—approaching the current limits of photonic integration and containing over 2,000 MZI units—conventional GPU-accelerated methods (e.g., using NVIDIA A100) require an estimated 1,000 hours (3.6×10⁶ seconds) to compute a solution, severely constraining practical deployment. In contrast, the MPIM-accelerated Ising approach accomplishes the same reconfiguration task in just 30 milliseconds, improving computational speed by nearly eight orders of magnitude and removing a major bottleneck for real-world applications of large-scale programmable PICs.

Looking ahead, further optimization will require representing more complex physical states of MZI units using additional Ising spins, in order to enable richer functional reconfiguration. However, this multibit spin representation significantly increases the demand on the optical memory capacity of MPIM hardware. Addressing this challenge will be a critical direction for future research.