Speeding up physics modeling through machine learning - Case study: Boltzmann Generators

Summary

Solving physics models is a need that has arisen for many decades. During this period, several critical issues have been resolved that impact a variety of domains in real life. This constant progress led to further investigation of physics problems, in greater detail and this procedure became more and more demanding. The same approach is met also in the domain of material science through molecular modeling, where their properties are examined and linked to their behavior at the level of atoms and molecules. Molecular modeling is a scientific field that gathers great research interest. It includes a variety of methodologies and techniques that cover a surprisingly large range regarding the properties of interest in each case. The corresponding computational schemes are based on physics and their application has great demands on computing resources and time.

These requirements increase in proportion to the complexity of the systems under study. This fact has led to research the use of various machine learning methods as part of exploring their applications. In molecular simulations, a small but characteristic control volume of the system is examined, accompanied by the characteristics of the component molecules and the interactions that develop between them. Then the behavior of the system is examined by using statistical physics and computational techniques. Also, the properties of the system are determined and linked to underlying molecular mechanisms. This research flow leads to understand in depth the operation of the specific system and therefore their dependence on several conditions can be determined, along with other factors that affect each system under investigation. In all cases, if the approach of examination is appropriate for a system depends on the correct determination of the model parameters and the selected methodology. The molecular simulations usually result in a large amount of raw data to be analyzed and managed.

This data usually includes configurations of the system, such as atoms' position and velocity in the considered control volume of the system. The quality of the results depends on their realism and probability. This thesis deals with the testing and evaluation of a recent approach for the creation of initial realistic molecular conformations through machine learning techniques, which differs from the classical construction effort to solve the problem. In specific, a neural network is implemented, which incorporates the Boltzmann generators method for the sampling of polyethylene molecules. This method has been initially applied to the generation of protein conformations where, unlike flexible polymers, their conformation depends on a relatively limited number of internal degrees of freedom. The model's training is extracted using Molecular Dynamics (MD) simulation. At last, a comparison is recorded against a classical Monte Carlo (MC) method of constructing molecular configurations already implemented.