Probleemstelling:

In modern society, the demand for energy is nearly insatiable. To keep up, we need to find new materials capable of generating, converting and transporting energy with high efficiency.[1,2] These materials are called energy materials and include applications such as photovoltaics, thermoelectrics and superconductors. The problem with these materials is that you cannot evaluate their performance without complex preprocessing and an advanced experimental setup. Given a random material, one cannot say in advance whether it will perform well or badly for the property of choice. Trying to find new materials thus becomes a costly and tedious process.

Initiatives such as the Materials Project[3] have tried to accelerate this search by creating large databases of properties from quantum simulations for both existing and hypothetical materials. For each material simulated, there are however millions of hypothetical materials that have not yet been explored. At the same time, the properties available in these databases are limited and many highly relevant experimental properties are so computationally expensive they are difficult to simulate even for a single material.

What if we could skip all that? We now have a new partner in our quest for new materials: deep learning. Deep learning is a subfield of machine learning specializing in extracting hidden patterns from large datasets and storing them in complex non-linear functions. Given enough training data, deep learning is able to calculate properties with similar performance to quantum mechanical methods, with a speed gain of several orders of magnitude[3,4]. This makes it possible to create quick, yet accurate, structure-property models which we can directly apply to existing databases to find new energy materials. 

Doelstelling:

In this thesis, we aim to find new energy materials by developing structure-property models using deep learning, based on existing property databases. The resulting models can then be applied to structural databases to find new candidate materials. One of the key new aspects in this thesis is that we will also include information from non-equilibrium structures. This means the model will learn how the properties change when the geometry changes, allowing us to both expand the applicability of our model to new geometries and optimize the geometry for the relevant properties. One way of achieving this is by including data from phonons.

Key points:

• Use and optimize deep learning models that link the crystal structure directly to properties such as energies, bandgaps and conductivities, but also dynamical properties such as phonons.
• Apply the developed models to new structural databases, including our in-house database of new quaternary materials, predicted to be stable and rank them for applications.
• Examine the ability of the models to directly optimize the crystal structure and find new materials.
• Gain understanding in the information considered important for the prediction of specific properties by the deep learning model and try to convert this to new physical insights.