Abstract
While the popularity of physics-informed neural networks (PINNs) is steadilyrising, to this date PINNs have not been successful in simulating dynamicalsystems whose solution exhibits multi-scale, chaotic or turbulent behavior. Inthis work we attribute this shortcoming to the inability of existing PINNsformulations to respect the spatio-temporal causal structure that is inherentto the evolution of physical systems. We argue that this is a fundamentallimitation and a key source of error that can ultimately steer PINN models toconverge towards erroneous solutions. We address this pathology by proposing asimple re-formulation of PINNs loss functions that can explicitly account forphysical causality during model training. We demonstrate that this simplemodification alone is enough to introduce significant accuracy improvements, aswell as a practical quantitative mechanism for assessing the convergence of aPINNs model. We provide state-of-the-art numerical results across a series ofbenchmarks for which existing PINNs formulations fail, including the chaoticLorenz system, the Kuramoto-Sivashinsky equation in the chaotic regime, and theNavier-Stokes equations in the turbulent regime. To the best of our knowledge,this is the first time that PINNs have been successful in simulating suchsystems, introducing new opportunities for their applicability to problems ofindustrial complexity.