シミュレーション科学の力による流体現象の解明を行うComputational Fluid Dynamics (CFD)とともに、データ科学の力を借りる Data-driven Fluid Dynamics (DFD) の進展も目覚ましい。 東京工業大学大西研究室では、CFDだけでなく、DFDによる環境流体研究の新たな展開を目指している。 このセミナーシリーズでは、関連分野の最新研究をおこなっている研究者に話題を提供していただく。
|
発表者: Youcef Belhadef, 1) Master student, Prof. Kai Schneider’s lab, Aix-Marseille University, CNRS, France 開催日時: 2026年4月16日（木）15:00-16:00 開催場所: 東工大大岡山キャンパス石川台9号館208室 (アクセス) |
|
|---|---|
| タイトル |
Physics-Informed Machine Learning for Plasma Turbulence Prediction: Neural Implicit Flow Combined with Wavelet Transform |
|
アブストラクト |
Predicting the spatio-temporal dynamics of edge plasma turbulence is essential for improving confinement in Tokamak fusion reactors. We present an approach combining Neural Implicit Flow (NIF) with discrete wavelet transform (DWT) to predict vorticity fields governed by the Hasegawa-Wakatani model. The NIF architecture decouples spatial complexity from parametric and temporal dependencies through a hypernetwork that dynamically generates the weights of a SIREN-based shape network. Wavelet preprocessing using Coiflets compresses high-resolution DNS data from 1024×1024 to 512×512 while preserving physically relevant coherent structures. We demonstrate that NIF-DWT achieves a mean SSIM above 0.90 and a PSNR gain of over 10 dB compared to the same architecture applied to raw data, while significantly reducing computational cost. The model is conditioned on the adiabaticity parameter, enabling predictions across multiple physical regimes without retraining. We compare our results against a ConvLSTM baseline and discuss the observed structure-amplitude decoupling, where the model accurately captures vortex geometry but shows variability in intensity prediction. |