Structure and properties of amorphous silicon from machine learning
Electronic fingerprints of structural transitions in dense Si © 2021 Volker Deringer / Springer Nature
A combination of machine learning models predicting the stability as well as the electronic properties of materials with the accuracy of quantum mechanics allowed to understand the structural transitions that silicon undergoes when compressed to tens of GPa. The study, performed by an international team from Oxford, Cambridge, the US Naval Research Laboratory and Ohio University, as well as EPFL's Laboratory of Computational Science and Modeling, has been published in Nature.
Machine learning is here to stay in materials modeling, and potentials that describe how atoms interact with each other, and determine the stability of structures at the nanoscale have already widely used. The next frontier involves predicting other properties, such as those associated with the electronic structure, with the same ease and accuracy. Recent work in the laboratory of Computational Science and Modeling have made this goal possible, and it is already possible to use these combined machine-learning models to address difficult questions concerning the behavior of complex materials in extreme conditions.
A paper recently published in Nature, provides an excellent example of how these simulation techniques can come together to solve the problem of how amorphous silicon transforms when compressed to 20GPa. Interatomic potentials based on machine learning reveal a sequence of structural transitions, from a low-pressur amorphous to a very high density amorphous phase, and then to a dense crystalline phase. A state-of-the-art model of the electron density of states allows elucidating how these transitions affect the conductivity of the sample.