Atomistic and Machine Learning Simulations for Nanoscale Thermal Transport
thesisposted on 26.07.2021, 21:34 authored by Prabudhya RoychowdhuryPrabudhya Roychowdhury
The recent decades have witnessed increased efforts to push the efficiency of energy systems beyond existing limits in order to keep pace with the rising global energy demands. Such efforts involve finding bulk materials and nanostructures with desired thermal properties such as thermal conductivity (k). For example, identifying high k materials is crucial in thermal management of vertically integrated circuits (ICs) and flexible nanoelectronics, which will power the next generation personal computing devices. On the opposite end of the spectrum, designing ultra-low k materials is essential for improving thermal barrier coatings in turbines and creating high performance thermoelectric (TE) devices for waste heat harvesting. In this dissertation, we identify nanostructures with such extreme thermal transport properties and explore the underlying phonon and photon transport mechanisms. Our approach follows two main avenues for evaluating potential candidates: (a) high fidelity atomistic simulations and (b) rapid machine learning-based property prediction and design optimization. The insight gained into the governing physics enables us to theoretically predict new materials for specific applications requiring high or low k, propose accelerated design optimization pathways which can significantly reduce design time, and advance the general understanding of energy transport in semiconductors and dielectric materials.
Bi2Te3, Sb2Te3 and nanostructures have long been the best TE materials due to their low κ at room temperatures. Despite this, computational studies such as molecular dynamics (MD) simulations on these important systems have been few, due to the lack of a suitable interatomic potential for Sb2Te3. We first develop interatomic potential parameters to predict thermal transport properties of bulk Sb2Te3. The parameters are fitted to a potential energy surface comprised of density functional theory (DFT) calculated lattice energies, and validated by comparing against experimental and DFT calculated lattice constants and phonon properties. We use the developed parameters in equilibrium MD simulations to calculate the thermal conductivity of bulk Sb2Te3 at different temperatures. A spectral analysis of the phonon transport is also performed, which reveals that 80% of the total cross-plane k is contributed by phonons with mean free paths (MFPs) between 3-100 nm.
We then use MD simulations to calculate phonon transport properties such as thermal conductance across Bi2Te3 and Sb2Te3 interface, which may account for the major part of the total thermal resistance in nanostructures. By comparing our MD results to an elastic scattering model, we find that inelastic phonon-phonon scattering processes at higher temperatures increases interfacial conductance by providing additional channels for energy transport. Finally, we calculate the thermal conductivities of Bi2Te3/Sb2Te3 superlattices (SLs) of varying period. The results show the characteristic minimum thermal conductivity, which is attributed to the competition between incoherent and coherent phonon transport regimes. Our MD simulations are the first fully predictive studies on this important TE system and pave the way for further exploration of nanostructures such as SLs with interface diffusion and random multilayers (RMLs).
The MD simulations described in the previous section provide high-fidelity data at a high computational cost. As such, manual intuition-based search methods using these simulations are not feasible for searching for low-probability-of-occurrence systems with extreme thermal conductivity. In view of this, we use machine learning (ML) techniques to accelerate and efficiently perform nanostructure design optimization within such large design spaces. First, we use a Genetic Algorithm (GA) based optimization method to efficiently search the design space of fixed length Si/Ge random multilayers (RMLs) for the structure with lowest k, which is found to be lower than the SL k by 33%. By comparing thermal conductivity and interface resistances between optimal and sub-optimal structures, we identify non-intuitive trends in design parameters such as average period and degree of randomness of layer thicknesses.
While machine learning (ML) has shown increasing effectiveness in optimizing materials properties under known physics, its application in discovering new physics remains challenging due to its interpolative nature. We demonstrate a general-purpose adaptive ML-accelerated search process that can discover unexpected lattice thermal conductivity (k) enhancement in aperiodic superlattices (SLs) as compared to periodic superlattices, with implications for thermal management of multilayer-based electronic devices. We use molecular dynamics simulations for high-fidelity calculations of k, along with a convolutional neural network (CNN) which can rapidly predict k for a large number of structures. To ensure accurate prediction for the target unknown SLs, we iteratively identify aperiodic SLs with structural features leading to locally enhanced thermal transport and include them as additional training data for the CNN. The identified structures exhibit increased coherent phonon transport owing to the presence of closely spaced interfaces.
We also demonstrate the application of ML in optimization of photonic multilayered structures with enhanced reflectivity to radiation heat flux, which is required for applications such as high temperature thermal barrier coatings (TBCs). We first perform a systematic variation of design parameters such as total thickness and average layer thickness of CeO2-MgO multilayers, and quantify their influence on the spectral and total reflectivity. The effect of randomization of layer thicknesses is also studied, which is found to increase the reflectivity due to localization of photons in certain spatial regions of the multilayer structure. Next, we employ a GA search method which can efficiently identify RML structures with reflectivity enhancements of ~22%, 20%, 20% and 10% over that obtained in randomly generated RML structures for total thicknesses of 5,10,20 and 30 microns respectively. We also calculate the spectral reflectivity and the field intensity distribution within the optimal and sub-optimal RML structures. We find that the electric field intensity can be significantly enhanced within certain spatial regions within the GA-optimized RMLs in comparison to non-optimized and periodic structures, which implies the high degree of randomness-induced photon localization leading to enhanced reflectivity in the GA-optimized structures.
In summary, our work advances the design or search for materials and nanostructures with targeted thermal transport properties such as low and high thermal conductivity and high reflectivity. The new insights provided into the underlying physics will guide the design of promising nanostructures for high efficiency energy systems.