DeePCK v1.0.0¶
- Author
Yuxiao Yi
- Contact
- Date
Oct 28, 2022
DeePCK is a Python package and aims at building alternative DNN models for Chemical Kinetics stiff Ordinary Differential Equations in numerical combustion simulation. DeePCK consists of three fundamental modules: Dataset, Model and Visualization. The source code of DeePCK will be publicly released on GitHub in the future. Further theoretical and technical details could be viewed in the paper: CombustFlame2022
The Dataset Module currently provide several sampling methods, such as adaptive zero dimensional
manifold (zero D), one dimensional flame (one D), multi-scale sampling (MS) and modified multi-scale sampling methods (MMS).
Besides, the label generating tool and data processing procedure are also included. All of them support massive parallelization via multiprocessing or MPI.
The Model Module deals with the framework design and training procedure of deep neural networks (DNNs), and the calling of DNNs for downstream tasks e.g. single-step prediction or temporal evolution. For instance, fig. 5 demostrates the zero-dimensional homogeous ignition process of CH4/air mixture simulated by Cantera and a well-trained DNN, respectively.
Fig 5 : Auto-ignition of CH4/air mixture.
Then the Visualization Module offers the interface to draw distribution and phase diagrams w.r.t typical chemical/thermodynamic datasets organized as \(T,P,Y\).
The model provided by DeePCK has coupled with stardand CFD codes to replace classical ODE solvers and shows promosing robustness and accuracy in high-dimensional flame simulation. We will continue to improve the corresponding source code package and documentation. If you have any issue or suggestion, feel free to contact the author via the email.
