Learning Across Dynamical Systems: Meta-Learning, Transfer Learning, and In-Context Identification

Organizers

Scope

System identification is usually carried out one system at a time: data are collected from a single plant and a model is fitted. Many applications instead involve families of similar systems, such as motors from the same production line, cells in a battery pack or vehicles in a fleet. Identifying each of them from scratch is inefficient, as it ignores what has been learned on the others. This track is about methods that learn across such classes of systems, so that a new member can be identified from little data and limited computational resources.

In transfer learning, a model pretrained on a class of systems is fine-tuned on limited data from a new one. Meta-learning includes this adaptation step in the training objective. Gradient-based methods learn initialisations from which a few updates suffice, and hypernetworks map a dataset directly to model weights. In-context methods skip the model update altogether: a sequence model reads a short input–output record of a system and directly predicts its response. Foundation models extend pretraining to heterogeneous time series from many domains.

Topics of interest

Theoretical, methodological and applied contributions are welcome. Topics include, but are not limited to:

• Meta-learning algorithms for identification: gradient-based methods, hypernetworks, in-context learning, amortised inference.
• Pretraining on synthetic data: design of simulated system classes and sim-to-real transfer.
• Foundation models for dynamical systems and time series.
• Physics-informed and grey-box meta-learning.
• Probabilistic and hierarchical Bayesian meta-learning, uncertainty quantification.
• Transfer learning and domain adaptation across plants and operating conditions.
• Meta-generalisation beyond the training distribution: distributionally robust optimisation, contrastive and invariance-based objectives.
• Sample complexity, identifiability and the relation to classical estimators.
• Self-tuning and predictive control with meta-learned models.
• State estimation, fault detection and diagnosis across system classes.
• Industrial applications, benchmarks and open datasets.

Submitting a contribution

Contributions are submitted through the IFAC PaperCept platform. Select Open Invited Track Paper and use the submission code cbk9x.

• Submission deadline: 2 November 2026
• Paper length: 4–8 pages at submission; final manuscripts are limited to 6 pages in the proceedings.
• Format: standard IFAC two-column, as specified in the IFAC Author Guide.

Submissions must be original and not under review elsewhere.

Other types of contribution

The procedure above is for regular track papers. To contribute a joint journal/conference paper, a dissemination paper or a discussion paper, please contact Marco Forgione.