The Dynamic Systems Lab (DSL) proudly participated in the International Conference on Machine Learning (ICML) 2025, held in Vancouver, Canada — one of the world’s premier gatherings for cutting-edge research in artificial intelligence, data science, and computational modelling.
This year, our group presented two papers that extend DSL’s mission to bridge dynamical systems theory with modern machine learning architectures:
🌀 Chaos Meets Attention: Transformers for Large-Scale Dynamical Prediction
This work introduces a transformer-based framework for learning and forecasting high-dimensional chaotic dynamics. By integrating attention mechanisms with the underlying structure of dynamical systems, the model captures long-range dependencies and multi-scale interactions that are traditionally challenging for recurrent or convolutional architectures. The findings shed light on how complex spatio-temporal patterns — from turbulent flows to climate models — can be learned efficiently through scalable, interpretable representations.
🧮 Tensor-Var: Variational Data Assimilation in Tensor Product Feature Space
This paper presents a new approach to variational data assimilation that leverages tensor product feature spaces to reconcile model dynamics with sparse, noisy observations. The method achieves efficient uncertainty propagation and improved convergence in large-scale systems, offering a new mathematical foundation for hybrid physics–AI data assimilation.
At ICML 2025, these contributions drew strong interest from researchers in machine learning for physical sciences, data-driven forecasting, and operator learning. The discussions reinforced DSL’s central research vision: developing representation, forecasting, control, and optimisation frameworks that unify data and physics across complex dynamic systems.
The lab extends sincere thanks to the ICML organising committee for an exceptional conference experience, and to the global research community for the stimulating exchange of ideas in Vancouver.
📄 Papers 1: “Chaos Meets Attention: Transformers for Large-Scale Dynamical Prediction”
📄 Papers 2: “Tensor-Var: Variational Data Assimilation in Tensor Product Feature Space”
📍 Conference: ICML 2025, 13-19 July, Vancouver, Canada
