Moiré Phonons in Twisted Bilayers of Transition Metal Dichalcogenides

Thanks to Prof. Nicholas Hine, Dr. Samuel Magorrian and Anas Siddiqui for taking me on for this project!

Introduction

This project successfully developed a foundation to study twisted bilayers of transition metal dichalcogenides (TMDs) using machine-learned interatomic potentials (MLIPs). A TMD is a hexagonal lattice of a metal $M$ and 2 atoms of a chalcogen $X$, denoted with symbol . TMDs are being heavily studied because of their electronic properties (bandgap), atomic size which make them extremely applicable to a wide range of devices including spintronics, optoelectronics, quantum computing technologies, medicinal technology and so on. TMDs are a class of 2D materials, i.e. crystal structures whose thickness are numbers of angstroms. Recent papers show promising results for twisted-bilayer TMDs, where we vertically stack TMDs and misalign their 2 layers by some twist angle $\theta$. This has led to the novel field of twistronics where we study how $\theta$ affects their electronic and mechanical properties.

Why did I decide to take on this project?

Quantum mechanics, programming and simulation all put together. I find it fascinating as to how physicists discover properties of materials. Most of it is done through Density Functional Theory (DFT) simulations with different potentials to try and extract the behaviour of materials. However, there is one issue - DFT simulations scale terribly with the number of atoms, and realistically the applications of TMDs will be in the thousands of atoms, which will take at least 1-2 weeks to run. Enter: MACHINE LEARNING. By training on a large dataset, this introduces the viability of huge speedups with minimal loss of accuracy. This is beneficial in multiple ways:
  1. Faster simulations means more research can be done.
  2. Faster simulations allow high performance computing resources to be used more efficiently, a good impact on the environment.
  3. Viability of ML can be pushed to other areas of research and allow us to access deeper physics - i.e. a positive feedback loop of the above.
Taking this project has been a 3-fold learning experience for me, to learn about DFT, phonons and ML in Physics.

What did I do?

  1. Acquire (not make, unfortunately) a ML model. The model I used was being trained by Anas. I did however learn how to make a small model, but nothing useful for this project.
  2. Generate twisted homobilayer $\text{MoS}_{2}$ (Molybdenum Disulfide) supercells of different sizes and twist angles.
  3. Run relaxation and phonon calculations on these supercells. These were run with a small GPU so took a while for a few hundred atoms.
  4. Scaling up to Molecular Dynamics software: LAMMPS. I parallelised this such that a 17862 atom system could be relaxed in 48 hours!
  5. Tried for phonon bandstructures using LAMMPS, but my lack of understanding of the parameters to do so hindered progress until time ran out. I did however, achieve a working method with phana.

What is next?

  • Further training on an improved model, and across heterobilayer TMD - this is being done brilliantly by Anas.
  • Producing phonon bandstructures on MD software to compare accuracy with DFT.