We combine quantum mechanical simulations with machine learning and optimization algorithms to computationally design materials with desired properties for various applications
Our main research pursuits are:
Through the portal of computer simulations we gain access to the vast configuration space of materials structure and composition. We can explore the uncharted territories of materials that have not been synthesized yet and predict their properties from first principles, based solely on the knowledge of their elemental composition and the laws of quantum mechanics. Since the Schrödinger equation can be solved exactly only for very small systems (=the hydrogen atom), we employ approximate methods within the framework of density functional theory (DFT) and many-body perturbation theory (MBPT) to apply quantum mechanics to systems, such as molecular crystals and interfaces, with up to several hundred atoms. The computational cost of quantum mechanical simulations increases rapidly with the accuracy of the method, the size of the system, and the number of trial structures sampled, therefore we run our calculations on some of the world’s most powerful supercomputers.
To navigate the configuration space and identify the most promising candidates, we use optimization algorithms. For example, genetic algorithms are guided to the most promising regions by the evolutionary principle of survival of the fittest. Machine learning (ML) uses statistical models based on “training data” to make predictions for new data points. We employ ML to accelerate predictions for materials properties and unveil hidden correlations in data generated by our simulations. We apply several types of ML algorithms for different purposes, such as optimization, classification, clustering, feature selection, sampling, and finding structure-property correlations. ML algorithms are integrated with quantum mechanical simulations in fully automated complex workflows.