Machine Learning to Construct Process-Structure-Property Relations of hBN Thin-Films

Hexagonal Boron Nitride (hBN) is a 2D material with high in-plane thermal conductivity and low out-of-plane thermal conductivity. Due to its anisotropic properties, hBN films have high potential for the thermal management of next-generation electronics. The thermal conductivity of hBN film depends on the processing conditions and its thickness. The construction of Process-Structure-Property (PSP) relations of hBN film can help in accelerating the synthesis of hBN films with desired thermal properties. Machine learning (ML) techniques can help in utilizing both measurements and simulations to develop reliable PSP relations to guide the synthesis process. In this presentation, we will discuss how Gaussian Process Regression (GPR) can be applied to experimental and simulation data to construct PSP relations. Pulsed laser deposition (PLD) is used for the extraction of experimental data, and molecular dynamics simulations is used for the extraction of simulation data. The PSP relations will be constructed using SEM images as structure data and out-of-plane thermal conductivity as property for various process conditions.<br/>A significant challenge is dimensionality mismatch which occurs from the large size of the structural data compared to process conditions and properties, e.g., SEM images are 2D images of a 3D structure, while process conditions or properties are typically 1D. Dimensional mismatch makes the ML model difficult to learn. To resolve this problem, we will apply two-point correlations and principal component analysis. Two-point correlations preserve spatial statistical information of the SEM images. This will allow original structural data from the samples to be converted into similar shapes, allowing the ML model to be trained efficiently. Applying principal component analysis will allow the structure data to be expressed as a linear combination of orthogonal vectors. The data will be represented by the matrix of coefficients of orthogonal vectors, which will be much smaller than an SEM image itself. Experimental data have high fidelity but low quantity, while simulation data can be produced in larger volume. Simulation data will be included with the experimental data during the training of ML model to accelerate PSP relations and the model will identify process conditions corresponding to high thermal conductivities accurately.