Stochastic PINNs solvers|| Multi-Resolution Independent Simulations for Turbulence || March 27, 2026

Описание: Speakers, institute & title
1) Prof. Paweł Przybyłowicz and Marcin Baranek, AGH University of Krakow, Poland, "Stochastic PINNs-deep learning solvers for SDEs"

Abstract: The talk presents a new methodology, called stochastic PINNs (StPINNs), for approximating sample paths of stochastic differential equations (SDEs) using artificial neural networks. The method is based on a Doss–Sussman transformation of the initial SDE into a random ordinary differential equation (RODE). The approach targets pathwise accuracy, highlights practical training strategies, and is demonstrated on benchmark SDE models. The talk is based on: https://arxiv.org/abs/2512.14258
_________________________________________________________________________
2) Prof. P. K. Yeung, Georgia Tech, "Exascale Computing and Multi-Resolution Independent Simulations Paradigm for the Study of Intermittency in Turbulence"

Abstract: Turbulence simulations in simplified geometries, including forced isotropic turbulence, have benefitted significantly from the advent of leadership-class GPU-based computing in the last few years. However, while advances in memory have enabled increasingly large problem sizes, strict time-stepping requirements in the Navier Stokes equations imply that, in practice, simulations at the highest resolution feasible at a given time are inevitably relatively short in time. In this talk we will discuss the use of a “Multi-Resolution Independent Simulations” (MRIS) approach which replaces a long simulation spanning multiple large eddy time scales by ensemble averaging over multiple short simulation segments, each of them having evolved by progressive grid refinement from lower-resolution snapshots. This approach is particularly suited to the study of small scale motions which evolve rapidly and likewise adjust quickly to changes of resolution in either space or time. MRIS has been used to obtain results for forced isotropic turbulence at grid resolution up to 327683 (or 35 trillion) grid points and Reynolds number comparable to many laboratory experiments. We also discuss the major elements of GPU acceleration for Fourier pseudo-spectral algorithms as well as particle tracking on the world’s first exascale computer (Frontier, at the DOE Oak Ridge Leadership Computing Facility).