Fluid flow prediction technology has been lately enhanced by ML-augmented and surrogate methods, and, increasingly need testing against challenging new datasets. Our competition invites all comers to test their method against a novel dataset of Navier-Stokes fluid flow prediction with unobserved latent factors. Competitor's methods will be tested against a hold-out test set, and measured in terms of time-normalised variance of the predictions on standard hardware.
Physics-informed neural networks (PINNs) are a type of machine learning methodology that incorporates physical laws and constraints into the training process. While they have shown tremendous potential across various domains, challenges remain in optimizing these physics-informed neural networks with conventional gradient-descent-based methodologies such as SGD. In preliminary investigations, evolutionary algorithms have shown potential as an effective alternative for PINN training. To further inspire advancements in evolutionary algorithms, and advance the possibilities of PINN models in modelling real-world complex problems, this competition is proposed centering on the development of evolutionary algorithms for the fast and effective training of PINN models as represented by 5 benchmark problems. These set of 5 benchmark problems were chosen to be diverse in type and for their relevance to real-world problems, comprising both ordinary and partial differential equations that describe phenomena in classical mechanics, heat and mass transfer, fluid dynamics and wave propagation (e.g. in acoustics).