Machine Learning for Elastic Strain Engineering

Elastic Strain Engineering

Elastic Strain Engineering (ESE) is an approach to controlling materials properties at nanoscale (or atomistic scale) through applying high strain or strain gradient to a material. ESE has many advanced applications in electronics, photonics, and mechanical engineering. Allowing for high strains and strain gradients greatly expands the materials parameter space being explored, handling which requires Big Data (machine learning) approaches of representation, exploration, and optimization of the respective physical properties such as the electronic band structure or the region of mechanical stability.

The focus of this project is on such machine learning approaches. The work requires close collaboration with Ju Li’s research group at MIT that will be responsible for computation of the relevant physical properties of materials.

Nanostructured materials can withstand much higher elastic strain without mechanical failure than their conventional counterparts, opening up a huge parameter space for rational engineering of material properties. We study inducing changes of technologically relevant properties in semiconductor crystals under high strains, including band gap and band structure. We use high-throughput ab initio calculations for the development of advanced machine learning modules to efficiently represent the dependence of material properties on the strain.

Band Structure Fitting and Strain Optimization

One of the main problems of the ESE project is to create a model for fitting the energy bands (mostly, the 4th and 5th ones) to be able to calculate the band gap and material properties such as whether it is a metal or semiconductor under the strain. At the moment, with data from ab initio calculations of ideal Si crystal, we fitted the ML model with a target accuracy of 0.007eV RMSE for the case of the diagonal matrix (no shear) and 0.0315eV for the full strain case. We created a dashboard where user can select the strain (from -5 to 10%) and see the band structure plot, see: .

With the help of such models and small amount of additional calculations, we are able to explore the space of strains which have the direct band gap, using recursive in-depth search in parameter space, basing on gradient-descent-like algorithms. It was done in case of ideal Ge crystal; although, the technique can be used for other materials as well.

Dynamical Matrix Model

We fit the dynamical matrix as a function of the strain and the wave vector. Through this, we can quickly determine if a given strain corresponds to a dynamically stable or unstable crystal. This, essentially, gives us the boundary of the physically relevant region in the strain space.