Machine Learning Algorithms and Molecular Dynamics Models for Predicting Nano-scale and Bulk Thermal Properties

Abstract

The engineering of nano-scale thermal transport mechanisms in a material can strongly influence its macro-scale behavior, with important implications for many thermal management applications. Atomistic computational methods in which the motion of individual atoms can be tracked offer a powerful means to study these mechanisms in circumstances for which performing experiments is challenging or impractical.

In this thesis, we first apply computational methods to study nano-scale heat transfer at strained interfaces between crystalline semiconductor structures. Such interfaces are important because of their applications in strained silicon transistors, thermoelectric materials, and lattice-mismatched epitaxial structures that are important for emerging applications (e.g., GaN on silicon). The strain at the interface disturbs the local lattice structure, which in turn alters phonon properties. We find that interfacial bonds between silicon and germanium atoms in a superlattice structure can introduce new vibrational modes in the system that reduce the interface thermal boundary resistance.

We then examine the relationships between bonding and thermal properties in the context of polymer chains, where inefficient inter-chain thermal coupling presents a bottleneck to macro-scale heat transfer. We consider various bonding/interaction types including covalent bonds, vdW and electrostatic interactions, and ionic interactions. We find that non-bonding interactions can have a significant impact on heat transfer in crosslinked polymers. For example, short crosslinkers can bring chains closer to each other and thereby increase inter-chain thermal conductance by non-bonding interactions. The understanding we gain by computational analysis is shown to resolve literature discrepancies regarding the effects of crosslinkers on heat transfer in polymers.

Finally, we develop a machine learning framework to compute the complex underlying relationships between a material’s basic physical properties (e.g., lattice structure), its local environmental properties (e.g., temperature), and its thermal properties. Our results show a ~five-fold reduction in simulation time versus current methods such as molecular dynamics or density functional theory. We also show how physical rules may be encoded in this and similar algorithms for materials property prediction such that the algorithm is not allowed to explore function spaces where physical rules do not hold. This inclusion of physical rules in the algorithm reduces the amount of data needed to train the algorithm, with broad applicability to machine learning of other material properties for which the feature size is large relative to available training data. These finding and models could pave the way toward more rapid design of engineered materials with desired thermal, mechanical, and electrical properties.