通过离线学习结合神经仿真实现可靠的粗粒度湍流模拟
流体动力学
2023-07-26 v1 大气与海洋物理
摘要
将未解析动态的机器学习(ML)模型集成到流体动力学数值模拟中,已被证明可提高粗分辨率模拟的精度。然而,当以纯离线模式训练时,将ML模型集成到数值格式中可能导致不稳定性。在一个二维准地转湍流系统背景下,我们证明在损失函数中加入一个将系统状态仿真至未来的附加网络,可产出捕获重要亚网格过程且稳定性更佳的离线训练ML模型。
引用
@article{arxiv.2307.13144,
title = {Reliable coarse-grained turbulent simulations through combined offline learning and neural emulation},
author = {Christian Pedersen and Laure Zanna and Joan Bruna and Pavel Perezhogin},
journal= {arXiv preprint arXiv:2307.13144},
year = {2023}
}
备注
Accepted after peer-review at the 1st workshop on Synergy of Scientific and Machine Learning Modeling, SynS & ML ICML, Honolulu, Hawaii, USA. July, 2023