Explainable Materials Informatics for Coherent Phonon Transport in 2D Heterostructures by Self-Learning Entropic Population Annealing

The exploration of coherent phonons demands flawless interfaces to circumvent phonon phase disruptions and inter-phonon interactions, a formidable challenge to achieve in practice. Van der Waals (vdWs) heterostructures, created through the precise layering of diverse two-dimensional materials, offer an optimal foundation for exploring coherent phonon transport because of their seamlessly integrated interfaces. Over the past decade, the comprehension of coherent phonon transport within superlattices has significantly advanced, propelled by the sophisticated field of materials informatics (MI) driven by machine learning techniques. Bayesian optimization, known for its efficiency in recommending structures with desirable properties, is a commonly adopted MI method. However, it’s a black-box method with the potential for converging to local minima and limited representation of the entire sample space. In contrast, entropic population annealing (SLEPA) extended from entropic sampling with a machine learning model by Li et al. [1], efficiently captures the density of states, enabling comprehensive thermal conductivity assessments for vdWs heterostructures while evaluating few candidates without missing optimal results.<br/>In this study, we obtain thermal conductivity distribution of graphene/WS₂ heterostructure in a candidate space of tens of thousands by combining SLEPA and atomistic Green’s functions. Though analysis, we identified two factors that have a significant negative impact on the thermal conductivities of heterostructures: the average distance of WS₂ from the center of the heterostructure and the presence of specific sequences, such as ‘1 0 0 1’, where ‘0’ represents graphene and ‘1’ represent WS₂. Moreover, by using mode-resolved atomistic Green’s function (AGF), the underlying mechanism under these two factors are investigated. The findings reveal that, as the average distance of WS₂ from the center of the heterostructure decreases, there is an increase in the transmission of high-frequency phonons at oblique angles. Additionally, an increase in the number of ‘1 0 0 1’ leads to higher transmission of low-frequency phonons that are normally incident. The combination of these two factors can suppress phonon transmission across a broad spectrum of frequencies and angles of incidence, leading to the optimized heterostructure with remarkably low thermal conductivity.<br/>[1] J. Li, J. Zhang, R. Tamura, K. Tsuda, <i>Self-learning entropic population annealing for interpretable materials design</i>, Digital Discovery 1, 295-302 (2022).