Multiscale Simulations of Complex Systems by Learning their Effective Dynamics

  • 2021-10-19 17:05:03
  • Pantelis R. Vlachas, Georgios Arampatzis, Caroline Uhler, Petros Koumoutsakos
  • 0

Abstract

Predictive simulations of complex systems are essential for applicationsranging from weather forecasting to drug design. The veracity of thesepredictions hinges on their capacity to capture the effective system dynamics.Massively parallel simulations predict the system dynamics by resolving allspatiotemporal scales, often at a cost that prevents experimentation whiletheir findings may not allow for generalisation. On the other hand reducedorder models are fast but limited by the frequently adopted linearization ofthe system dynamics and/or the utilization of heuristic closures. Here wepresent a novel systematic framework that bridges large scale simulations andreduced order models to Learn the Effective Dynamics (LED) of diverse complexsystems. The framework forms algorithmic alloys between non-linear machinelearning algorithms and the Equation-Free approach for modeling complexsystems. LED deploys autoencoders to formulate a mapping between fine andcoarse-grained representations and evolves the latent space dynamics usingrecurrent neural networks. The algorithm is validated on benchmark problems andwe find that it outperforms state of the art reduced order models in terms ofpredictability and large scale simulations in terms of cost. LED is applicableto systems ranging from chemistry to fluid mechanics and reduces thecomputational effort by up to two orders of magnitude while maintaining theprediction accuracy of the full system dynamics. We argue that LED provides anovel potent modality for the accurate prediction of complex systems.

 

Quick Read (beta)

loading the full paper ...