Recent decades have witnessed significant advancements in the field of system identification. By analyzing input-output data and extracting mathematical models, system identification facilitates the characterization and prediction of dynamics, aiding in the subsequent controller design, optimization, and decision-making procedures. In traditional system identification, the focus is on learning a model of a specific system through pre-existing physical knowledge and measured input-output trajectories. This typical workflow is inextricably linked to supervised machine learning, i.e., discovering mathematical relationships in the data. One of the main strengths of machine learning approaches is that they can efficiently handle high-complexity function estimation problems although their application in the dynamic setting often lacks statistical guarantees such as consistency. While system identification tools rooted in statistics have been developed under various efficiency and statistical guarantees, they often have limitations in terms of the complexity of the estimation problems they can handle.

Hence, we recognize that fusing system identification and machine learning presents a remarkable opportunity. This fusion enables traditional system identification to efficiently handle nonlinear systems and extend machine learning to be applicable to commonly encountered complex dynamical systems in real life. However, the challenges revolving around fusing machine learning with system identification, as evidenced by current results, primarily center on two aspects: i) lacking physical interpretation; ii) leveraging knowledge accumulated across related systems and tasks. We also focus on the opportunities that this opens in terms of “new” problems that have appeared: transfer learning, learning dynamics during reinforcement learning and model predictive control, etc.

The workshop offers a repertoire of the most recent theoretical and practical developments of learning dynamical systems from data, both from the more classical system identification point of view as well as from a machine learning perspective. We also aim to cover a wide diversity of application domains, ranging from classical systems and control, to robotics, to general scientific learning. The key topics include physics-informed learning, meta-learning, and deep neural networks, with applications in building energy systems, electromechanical systems, mechatronic systems, etc. Beyond the presentations, ample time will be provided for discussion together with the audience. Furthermore, the workshop seeks to foster valuable interdisciplinary dialogue and collaboration between academia and industry.

This workshop is tailored for graduate students and researchers interested in system identification and machine learning, with particular challenges in designing high-performance identification and learning algorithms in real-world scenarios. Attendees are expected to have some basic technical background in system identification and machine learning, as the content is designed to meet the needs of graduate-level researchers both from academia and industry, ensuring that everyone can derive valuable insights from participation.