PHLieNet: Hypernetwork-Based Forecasting
Led by Pantelis Vlachas, in collaboration with Konstantinos Vlachas and Eleni Chatzi, this work forming part of the TURING Horizon project, presents a new hypernetwork-based approach for adaptive and generalisable forecasting in complex parametric dynamical systems.
The paper, entitled “Beyond Static Models: Hypernetworks for Adaptive and Generalizable Forecasting in Complex Parametric Dynamical Systems,” has been published in Mechanical Systems and Signal Processing. It is authored by Pantelis R. Vlachas, Konstantinos Vlachas and Eleni Chatzi from the ETH Zürich TURING team.
The study introduces PHLieNet, a Parametric Hypernetwork for Learning Interpolated Networks. Conventional approaches often require a separate forecasting model to be trained for each parameter configuration. In contrast, PHLieNet learns to interpolate directly within the space of forecasting models, allowing a unified framework to adapt across different dynamical regimes and generalise to previously unseen parameter settings.The proposed methodology demonstrates strong performance in short-term forecasting while also accurately representing the long-term behaviour of complex dynamical systems. The work contributes to TURING’s broader objective of developing robust, adaptive and generalisable artificial intelligence methods for modelling and predicting complex physical processes.The publication marks the first journal paper produced within the TURING project.