Using machine learning to study rarefied gases

April 24, 2023

Shahin Mohammad Nejad defended his PhD thesis at the department of Mechanical Engineering on April 6th.

Shahin Mohammad Nejad.

An understanding of rarefied gas dynamics has practical applications in numerous industrial and environmental areas, such as the semiconductor industry, micro- and nano-fluidics, aerospace engineering, nuclear reactor safety, heterogeneous catalysis, and seawater desalination. For these applications, a fundamental understanding of the physics of gas flow is crucial, which, at the design and fabrication stages in these applications, can be achieved using numerical tools that are selected based on the degree of rarefaction in the system and the computational power that can be afforded. For his PhD thesis, Shahin Mohammad Nejad used machine learning approaches to create a new model to study gas dynamics.

In many engineering systems, above a certain degree of rarefaction, important transport properties, such as shear stress, pressure drop, and heat flux, cannot be predicted using continuum flow and heat transfer models.

Under such circumstances, Boltzmann transport equations are typically applied to describe the flow field properties, which can be solved using different simulation techniques, such as Direct Simulation Monte Carlo (DSMC), Method of Moment (MoM), and Lattice Boltzmann Methods (LBM). For these techniques, an essential element is an accurate description of boundary conditions, which is the key factor towards achieving reliable simulation results.

Gas-surface interactions

Due to the complexity of the microscopic interaction between gas molecules and boundary surfaces, it is difficult to develop an analytically tractable and general model to describe gas-surface interactions.

To circumvent such complexity, many researchers combine their theoretical understanding of gas-surface interactions with either experimental measurements or molecular dynamics (MD) simulations and propose empirical boundary models with different levels of sophistication that can be used in various rarefied gas flow systems. However, such empirical boundary models in the case of systems with highly nonequilibrium or complex gas flow conditions cannot completely describe the gas-surface interactions.

MD simulations are viewed as a powerful computational tool to study gas鈥搒urface interactions at the atomistic level. In fact, through the use of MD simulations, it is possible to track individual molecules under a wide range of conditions and achieve a detailed understanding of momentum and energy exchange mechanism at gas鈥� solid surface interfaces. However, even with fast computers, it becomes very difficult to use a full MD approach to model gas鈥搒olid interactions in a physical system due to the excessive computational costs.

Machine learning

In his work, Shahin Mohammad Nejad constructed a generalized gas-surface interaction model based on insight gained from MD simulations and a Gaussian Mixture (GM) approach, which is an unsupervised machine learning approach.

The performance of the proposed boundary model was examined against the existing empirical gas-surface interaction models in different benchmarking systems. Mohammad Nejad鈥檚 research show that his new model outperforms many previous models.

Title of PhD thesis: . Supervisors: David Smeulders, Arjan Frijns, and Silvia Gaastra-Nedea.