Workshop: AI for Science (AI4S) in Context of Optical Fiber Communications

The emerging paradigm of AI for Science (AI4S) is fundamentally reshaping research across disciplines. Optical fiber communications, with its intricate nonlinear dynamics and multi-scale design challenges, stands as a prime beneficiary and testing ground. As optical systems approach fundamental capacity limits, the community faces two pressing and interconnected bottlenecks.

First, purely data-driven machine learning models, however powerful, are data-hungry and lack inherent respect for physical laws, limiting their reliability in extrapolation and their interpretability to domain scientists. Second, even when such models are developed, they remain largely confined to isolated simulation environments, requiring substantial manual effort to translate insights into experimental validation or deploy them within autonomous engineering workflows.

This workshop focuses on AI-driven scientific discovery, computation, and validation in the context of optical fiber communications, exploring how advanced AI techniques can enable automated scientific research and reliable engineering implementation.

The workshop is centered around two interconnected technical themes, both in the context of optical communications.

Theme 1: Data-Knowledge Hybrid Driven AI

This session covers methodologies for embedding physical laws such as partial differential equations (PDEs), symmetry constraints, and conservation principles directly into the architectures and loss functions of neural networks and Gaussian processes.

Key Topics

• High predictive accuracy with sparse data, model interpretability, and the application of "Learnable DSP" architectures to mitigate fiber impairments.
• A universal framework for solving nonlinear dynamics in fiber optics, including forward, inverse, and backward problems.
• Advanced neural network architecture: a hybrid data-driven and physics-informed design.

Theme 2: From AI Tools to Autonomous Scientific Agents

This session focuses on the development of LLM-powered multi-agent systems, self-driving laboratories, and scientific research robots, engineered to automate end-to-end scientific and engineering workflows.

Key Topics

• AI-enhanced scientific discovery for novel hypotheses, physical laws, and governing equations.
• Scientific agents for autonomous simulation construction, experiment validation, and result analysis.

Interactive Panel Session

The workshop will conclude with a structured panel discussion. Attendees will have the opportunity to engage with speakers via live digital polling on key industry questions:

1. What are the deployment timelines for replacing numerical split-step methods with physics-informed neural networks in commercial planning tools?
2. What are the primary standardization and safety bottlenecks preventing the integration of autonomous agents into physical optical laboratories?
3. How far can optical agents automate device, link, and network workflows? Where should human experts remain in the loop?
4. Which scientific, theoretical, or engineering design approaches are most likely to deliver near-term breakthroughs with advanced AI agents?
5. As LLM-powered multi-agent systems become capable of autonomously designing simulations and running experiments, how should we redefine the role of the human researcher—are we moving toward curators of AI reasoning or verifiers of AI outputs, and what new skills will that require?