论文标题
部分可观测时空混沌系统的无模型预测
Machine Learning-based models in particle-in-cell codes for advanced physics extensions
论文作者
论文摘要
储层计算是预测湍流的有力工具，其简单的架构具有处理大型系统的计算效率。然而，其实现通常需要完整的状态向量测量和系统非线性知识。我们使用非线性投影函数将系统测量扩展到高维空间，然后将其输入到储层中以获得预测。我们展示了这种储层计算网络在时空混沌系统上的应用，该系统模拟了湍流的若干特征。我们表明，使用径向基函数作为非线性投影器，即使只有部分观测并且不知道控制方程，也能稳健地捕捉复杂的系统非线性。最后，我们表明，当测量稀疏、不完整且带有噪声，甚至控制方程变得不准确时，我们的网络仍然可以产生相当准确的预测，从而为实际湍流系统的无模型预测铺平了道路。
In this paper we propose a methodology for the efficient implementation of Machine Learning (ML)-based methods in particle-in-cell (PIC) codes, with a focus on Monte-Carlo or statistical extensions to the PIC algorithm. The presented approach allows for neural networks to be developed in a Python environment, where advanced ML tools are readily available to proficiently train and test them. Those models are then efficiently deployed within highly-scalable and fully parallelized PIC simulations during runtime. We demonstrate this methodology with a proof-of-concept implementation within the PIC code OSIRIS, where a fully-connected neural network is used to replace a section of a Compton scattering module. We demonstrate that the ML-based method reproduces the results obtained with the conventional method and achieves better computational performance. These results offer a promising avenue for future applications of ML-based methods in PIC, particularly for physics extensions where an ML-based approach can provide a higher performance increase.
