Digital twin technology has emerged as a transformative approach to bridging the gap between the physical and digital worlds, enabling unprecedented levels of intelligence, automation, and optimization in complex systems. In the context of optical communication networks, the rapid growth of data traffic driven by new applications such as Internet of Things (IoT), artificial intelligence (AI), cloud computing, and ultra-high-definition video imposes stringent requirements on network capacity, latency, reliability, and flexibility. Traditional optical networks, while benefiting from advances in coherent transmission, elastic optical networks, and software-defined networking, still face challenges such as conservative link configurations, static control mechanisms, and resource over-provisioning during fault recovery. These limitations hinder the efficient utilization of spectrum resources and the realization of intelligent, dynamic network management. Digital twin technology offers a solution by enabling real-time, high-fidelity mapping of the physical network into a digital word, supporting accurate perception, prediction, and intelligent decision-making throughout the entire network lifecycle. This paradigm shift is essential for evolving from passive maintenance to proactive, autonomous operation in next-generation optical networks.

Significant advances have been made in the development of digital twin technologies for intelligent optical networks, particularly in two core areas: high-precision physical-layer modeling and data-driven self-learning. For physical-layer modeling, a modular approach is widely adopted, where each network element—such as fibers, erbium-doped fiber amplifiers (EDFAs), Raman amplifiers (RAs), and wavelength selective switches (WSSs)—is independently modeled and then integrated to construct an end-to-end digital representation of the optical link (Fig. 1). Traditional white-box models are based on physical principles. For instance, the Manakov equation and the Gaussian noise (GN) model can be utilized to model fiber nonlinearity. Ordinary differential equations (ODEs) can be used to model EDFAs and RAs. Though these white-box models provide strong interpretability, they are often constrained by computational complexity and idealized assumptions. To overcome the limitations of white-box models, black-box models leveraging machine learning techniques such as neural networks have been introduced, achieving high prediction accuracy and computational efficiency while closely matching the results of conventional simulations (Figs. 3 and 7). Furthermore, grey-box models that combine physical knowledge with data-driven learning have demonstrated improved accuracy, robustness, and generalizability by integrating the strengths of both white-box and black-box approaches. This modeling paradigm has been successfully applied to the characterization of fibers, EDFAs, and RAs, where advanced machine learning methods such as active learning significantly reduce the required training data without sacrificing accuracy (Figs. 5, 8, and 9). Similar strategies have been extended to the modeling of filtering penalty, further enhancing the accuracy of optical network digital twins.

After developing accurate physical models, data-driven self-learning becomes an essential technique for the deployment and continuous evolution of these models. This approach enables digital twins to autonomously adapt to dynamic network environments and maintain high modeling accuracy over time. The self-learning process primarily involves three key research aspects: telemetry, model inaccuracy correction, and input parameter uncertainty refinement. First, telemetry-driven data acquisition focuses on the real-time collection and integration of multi-source data from both the optical and digital domains. In addition to deploying devices such as optical power meters within optical links for monitoring purposes, acquiring real-time data from the digital signal processor (DSP) at the receiver has become a crucial approach. This method enables the real-time perception of multiple key metrics, including optical power evolution, nonlinear interference (Fig. 10), and filtering impairments (Fig. 11), thus providing a solid data foundation for the effective operation of subsequent self-learning algorithms. Second, in terms of model inaccuracy correction, techniques such as transfer learning and active learning have demonstrated the ability to significantly enhance the generalization and adaptability of models in real-world scenarios. Characterized by small sample sizes, the data dependency and training costs associated with large-scale deployment of digital twin systems can be reduced (Figs.12 and 13). Third, input parameter refinement techniques further formulate the uncertainty in input parameters as an optimization problem. By applying suitable optimization algorithms and various refining paradigms, uncertainties in the input parameters can be reduced, thereby reducing the impact of measurement noise and incomplete information on model performance (Figs. 14 and 15).

Digital twin technology is revolutionizing the management and optimization of optical communication networks by providing a dynamic, high-fidelity virtual representation of the physical infrastructure. The combination of physical-layer modeling and data-driven self-learning significantly enhances the accuracy, adaptability, and intelligence of network operations. Future research will focus on improving model scalability, data ecosystem integration, and trustworthiness. Challenges such as data sparsity, cross-domain generalization, and real-time deployment have yet to be solved. In the future, addressing these challenges through joint efforts among academia, industry, and research organizations will be essential for realizing the paradigm shift from passive maintenance to proactive prediction in optical network operations, thereby laying a solid foundation for the next generation of intelligent communication infrastructure.