Researchers at The Massachusetts Institute of Technology (MIT) have established a proposed method which merges machine learning with first-principles calculations to help in managing the computational complexities required in understanding the thermal conductivity of semiconductors, specifically focusing on diamonds. The diamond, known for its exceptional thermal conductivity, has several factors that complicate the conventional understanding of how its lattice thermal conductivity can be modulated through reversible elastic strain (ESE).
Traditionally, first-principles calculations have been utilized to comprehend the phonon band structure and related properties. However, these methods can be computationally expensive and may not be suitable for real-time computation. In the new approach, neural networks are deployed to acknowledge and make use of the structured relationship between band dispersion and strain.
The methodology consists of first, comparing computational results against experimental values for undeformed diamonds. About 15,000 strain points are collected using Latin-Hypercube sampling and then input into ab initio calculations to get the properties for each deformed structure. DFT simulations are used for structure relaxation and use the Green-Lagrangian strain measure. Phonon calculations are based on density functional perturbation theory (DFPT). Machine learning models, such as fully connected and convolutional neural networks, are then trained to make predictions about phonon stability, DOS, and band structures for a range of strain states.
The efficiency of the models is enhanced through strategic data sampling and active learning cycles. In addition, to compute a diamond’s thermal conductivity, molecular dynamics (MD) simulations are used. This provides a qualitative validation of the observed trends.
The innovative approach proposed in the research contributes significantly to understanding and modulating the thermal conductivity of diamonds through reversible elastic strain. By using machine learning models trained on first-principles calculations, properties of strained diamond structures, including phonon stability and related properties, can be predicted. This method brings forth a computationally efficient path to investigate the complex relationship between strain and thermal conductivity, unlocking possibilities for customizing the performance of devices and optimizing figure-of-merit in semiconductors.
This groundbreaking research heralds a new era in material science, offering a guide for the fine-tuning of specific material properties through the application of machine learning. It has the potential to revolutionize our understanding of diamonds while also demonstrating the broader implications of these methods for the study of semiconductors and other materials.
