Physics-Informed Learning for Modeling, Optimization, and Control of Cyber-Physical Systems

Machine learning is rapidly transforming how cyber-physical systems (CPS) and Internet-of-Things (IoT) infrastructures are modeled, optimized, and controlled. Modern CPS increasingly combine physics-based models, dense sensing, data-driven components, and networked decision loops operating at scale. This evolution exposes a gap between purely black-box ML approaches, which can be flexible but untrustworthy, and classical model-based engineering approaches, which offer guarantees but may lack adaptability and scalability. Physics-informed machine learning (PIML) provides a principled bridge between these worlds by embedding physical laws, structural priors, and engineering constraints directly into ML architectures, loss functions, and training pipelines. This leads to models and controllers that are more robust, data-efficient, interpretable, and suitable for safety-critical CPS/IoT domains such as smart buildings, transportation systems, robotics, energy grids, and industrial IoT processes. This half-day tutorial introduces the CPS-IoT community to the latest methods in physics-informed learning for modeling, optimization, and control. We provide foundational lectures on PIML concepts, differentiable programming, and physics-aware neural architectures as well as hands-on coding demonstrations using open-source toolchains such as Neuromancer and differentiable solvers in PyTorch, followed by supervised exercises that participants can run on their own laptops. The tutorial is designed for a broad audience ranging from graduate students to industry practitioners seeking trustworthy and data-efficient ML techniques for dynamical CPS. 

Outcomes: Participants will gain a working understanding of how to embed structural and physical knowledge into ML-based models and controllers for CPS. They will learn practical tools and workflows for implementing physics-informed neural networks, differentiable predictive control, and domain-informed system identification. They will also receive tutorial notebooks, reproducible examples, and reference implementations for their own research and teaching. The environment will encourage interactive discussion of open challenges, emerging research questions, and promising application domains.

Intended Audience: The intended audience includes graduate students, early-career researchers, and experienced practitioners in CPS, IoT, controls, machine learning, and embedded systems. The tutorial assumes only basic familiarity with control or machine learning concepts and is designed to be accessible and practical. It is particularly relevant to researchers working on digital twins, intelligent infrastructure, robotics, transportation systems, smart buildings, energy systems, industrial IoT, and other domains where data-driven and physics-based methods must coexist. All participants will have access to a website containing slides, code, and installation instructions prior to the event, ensuring seamless engagement during the hands-on sessions.