ULTRA-FAST DSMC BASED ON EXPLAINABLE AI FOR ALL FLOW REGIMES INCLUDING RAREFIED HYPERSONICS
Abstract
The direct simulation Monte Carlo (DSMC) method directly simulates the molecular behavior of gases by decomposing the motion of the particles into deterministic movement and stochastic collision with the assumption of one simulated particle representing a large number of real particles. The method has evolved into a primary workhorse to computationally solve the Boltzmann kinetic equation (BKE) and is routinely being applied to various flow problems of scientific and technological interest including rarefied hypersonic gas flows. However, the DSMC method suffers from very high computational costs, especially in the regime near the continuum limit and in three dimensional flow problems. In addition, it is very difficult to decipher the stochastic simulation process from the DSMC results, because the DSMC method is a pure simulation method. To overcome these shortcomings, this AFOSR Grant application aims to develop an innovative method for combining artificial neural network algorithms in AI suitable for highly nonlinear hypersonic flow problems with new interpretations of DSMC results in the context of constitutive modeling. The proposed method is expected to make DSMC ultra-fast, comparable to the computational cost of conventional CFD methods, and to provide multi-dimensional constitutive relationships that enable theoretical descriptions of physical phenomena for all flow regimes. The method is based on three key ideas. First, BKE, DSMC, and CFD all share the same conservation laws, but differ only in their constitutive relations. Second, constitutive modeling of gas flows described by DSMC is possible via decomposition of DSMC data and machine/deep learning. Lastly, conservative CFD methods in conjunction with DSMC constitutive models are equivalent to DSMC and their computational cost will be comparable to that of conventional CFD methods. The major outcomes are the core methodologies, validation study results, training data generation, and learning algorithms, which will be useful to rarefied hypersonics and AI communities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 20, 2023
- Source ID
- FA23862214051
Entities
People
- Rho Shin Myong
Organizations
- Air Force Office of Scientific Research
- United States Air Force