Fast Prediction of Multi-Mode Fiber Propagation Achieves Real-Time Results
Summarize this article with:
Understanding how ultrashort pulses of light travel through multimode optical fibres presents a significant challenge, as the process involves complex nonlinear interactions, and traditional computer simulations are often too slow for practical applications.
Dinesh Kumar Murugan and Nithyanandan Kanagaraj, both from the Indian Institute of Technology, Hyderabad, address this problem by developing a new approach to modelling light propagation. Their work introduces a powerful machine learning framework that not only predicts how light evolves within these fibres, but also rapidly reconstructs the original input light field from measured outputs, achieving a bidirectional capability. This innovative architecture functions as a fast and accurate substitute for conventional simulations, promising substantial speedups and enabling real-time beam diagnostics, the design of complex light patterns, and precise control over light’s behaviour in optical fibres and potentially other wave-based systems. In graded-index multimode fibers, nonlinear behaviour arises from several physical processes. Although conventional numerical solvers can reproduce this behaviour with high fidelity, their computational cost restricts real-time prediction, rapid parameter exploration, experimental feedback, and crucially, inverse retrieval of input fields from measured outputs. This work introduces an operator learning framework that learns both the forward and inverse propagation operators within a single, unified architecture. By combining spectral filters for spatio-temporal representations with Fourier-embedded conditioning on physical parameters, the model functions as a fast surrogate capable of efficiently addressing these challenges.
Neural Networks Model Multimode Fiber Optics Researchers are increasingly using neural networks to model and understand the complex behaviour of light travelling through multimode optical fibers. These fibers, which allow multiple light paths, offer increased bandwidth but present challenges in controlling light propagation and mitigating distortions. This work demonstrates how machine learning overcomes limitations of traditional simulation methods, enabling faster and more accurate predictions of light behaviour within these systems. Several machine learning techniques are proving valuable in this field, including recurrent and convolutional neural networks. Physics-informed neural networks combine machine learning with established physical laws, improving accuracy and generalizability. Fourier neural operators and DeepONet further enhance efficiency by leveraging mathematical transformations to solve complex equations.
This research focuses on developing methods for simulating fiber optic systems, processing the resulting signals, and optimizing system parameters. Data preprocessing and augmentation techniques improve the quality and quantity of data used to train these machine learning models, gaining new insights into the behaviour of light in complex fiber systems and developing innovative solutions for optical communication and sensing.
Operator Learning Predicts Pulse Propagation in Fibers Scientists have achieved a significant breakthrough in computational efficiency by developing a novel operator learning framework. This framework accurately predicts and reconstructs how ultrashort pulses of light propagate within graded-index multimode fibers. The system learns both the forward and inverse processes of light propagation within a single architecture, enabling rapid prediction of complex field evolution and reconstruction of input fields from measured outputs. This represents a substantial speedup compared to conventional numerical solvers, paving the way for real-time beam diagnostics and data-driven design of complex input fields. The research team generated a comprehensive dataset using detailed (3+1)D simulations based on the Generalized Nonlinear Schrödinger Equation, accurately modelling the behaviour of light within the fiber. These simulations, conducted over approximately 5.55mm, captured the intricate interplay of diffraction, chromatic dispersion, and nonlinear effects. Input pulses, consisting of a short duration Gaussian envelope combined with multiple fiber modes, varied in peak power, spanning a wide range of propagation regimes. At each step, the intensity of light was recorded, forming the basis for training the operator learning model. To enhance learning, each simulation was paired with its reversed counterpart, allowing the neural operator to learn both forward and inverse operations simultaneously. A binary indicator distinguished between forward and inverse modes, effectively doubling the dataset size. Data preparation involved scaling all values to improve numerical stability and ensure comparable contributions from each input parameter. The resulting model demonstrates a significant advancement in the ability to model and control light propagation in complex fiber systems. Fast Prediction of Pulse Propagation in Fibers This research presents a novel operator learning framework capable of accurately predicting and reconstructing ultrashort pulse propagation within multimode optical fibers. By combining spectral filters with conditioning on physical parameters, the team developed a model that functions as a fast surrogate for complex field evolution, achieving substantial speedups compared to conventional numerical solvers. This bidirectional architecture learns both forward and inverse propagation operators within a unified system, representing one of the first demonstrations of its kind applied to this specific problem. The resulting model enables rapid prediction of field behaviour, paving the way for real-time beam diagnostics, data-driven design of complex input fields, and closed-loop control of spatio-temporal dynamics. While the current work focuses on optical fibers, the authors suggest the framework’s potential applicability extends to a wide range of wave systems exhibiting similar nonlinear and dispersive effects. Future work will focus on improving the model’s generalizability and robustness, and exploring the physical meaning of the learned features to gain deeper insights into fiber design and optimization. 👉 More information 🗞 Bidirectional Fourier-Enhanced Deep Operator Network for Spatio-Temporal Propagation in Multi-Mode Fibers 🧠 ArXiv: https://arxiv.org/abs/2512.15474 Tags: