A Differentiable Hydrodynamic Simulator for Flow Design, Inference, and Learning
John Jairo Molina Lopez(Kyoto University / Department of Chemical Engineering)

We will develop an end-to-end differentiable direct numerical simulation (DNS) method for solving particle/fluid design, inference, and learning problems. For this, we will adopt the Smooth Profile Method, which allows us to directly solve the Navier-Stokes equation, while efficiently and accurately accounting for fluid-particle interactions. The method will be implemented in JAX/Python, which provides Automatic Differentiation support on multi-GPU systems. This will allow us to (1) solve large-scale flow optimization problems (e.g., optimizing for the fluid/particle properties), as well as (2) incorporate the DNS method within existing Machine-Learning frameworks. This will pave the way towards “designer” soft-matter systems, as well as provide a valuable tool to analyze experimental data.

研究紹介ポスター／最終報告書