In this paper we enhance the well-known fifth order WENO shock-capturing scheme by using deep learning techniques. This fine-tuning of an existing algorithm is implemented by training a rather small neural network to modify the smoothness indicators of the WENO scheme in order to improve the numerical results especially at discontinuities. In our approach no further post-processing is needed to ensure the consistency of the method, which simplifies the method and increases the effect of the neural network. Moreover, the convergence of the resulting scheme can be theoretically proven. We demonstrate our findings with the inviscid Burgers' equation, the Buckley-Leverett equation and the 1-D Euler equations of gas dynamics. Hereby we investigate the classical Sod problem and the Lax problem and show that our novel method outperforms the classical fifth order WENO schemes in simulations where the numerical solution is too diffusive or tends to overshoot at shocks.
翻译:在本文中,我们通过运用深层学习技术强化了众所周知的第五顺序WENO冲击冲击摄取计划。 对现有算法的微调是通过训练一个相当小的神经网络来实施, 以修改WENO计划的光滑性指标, 从而改进数字结果, 特别是在不连续的情况下。 在我们的方法中, 不需要再进行后处理来确保方法的一致性, 这种方法简化了方法, 并增加了神经网络的效果。 此外, 由此产生的方法的趋同性可以在理论上得到证明。 我们用隐蔽的Burgers方程式、 巴克利- 莱韦莱特方程式和气体动态的1D Euler方程式来展示我们的调查结果。 我们在这里调查典型的SOD问题和Lax问题, 并表明我们的新方法在模拟中超过了传统的第五顺序(WENO)方法, 因为数字解决方案过于分散或往往在冲击中过度解决。