Predicting large-scale behaviours of active matter models & graph networks

Dr Xietao Wang-Lin will be presenting Predicting large-scale behaviours of active matter models & graph networks as part of the SciML series
About this Event

A hybrid event held by the SciML Community at Leeds Institute for Data Analytics (LIDA)

Recent advancements in neural surrogate modelling for particle systems have achieved impressive results in engineering benchmarks, achieving large speed-ups while maintaining accuracy. However, zero-shot generalisation of the neural surrogate to much larger particle counts, rigid geometries and long-time horizons remain less studied. In this talk, we study at the application of neural surrogates trained on low particle counts to model large-scale phenomena exhibited in active matter models. First, we adapt the graph neural network-based surrogates to simulate first-order active matter systems and show that the surrogate is able to predict resonance in systems with lattice obstacles. Then, we show that the trained models are capable of zero-shot generalisation, predicting emerging behaviour such as synchronisation in the Vicsek model and clustering in system of active Brownian particles, but struggle to maintain proper particle packing. We believe this provides future work a better proxy for performance of neural surrogate models.

Biography: Xietao Wang Lin is a PhD student in the Mathematics for Real-World Systems Centre for Doctoral Training (MathSys CDT) at the University of Warwick, advised by Prof. Tom Montenegro-Johnson. His current research focuses on accelerating agent-based simulations using graph neural networks. Prior to starting his PhD, he earned his MSc degree from the University of Warwick, where his research focused on Bayesian optimisation and physics-informed neural networks.