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] These materials are called energy materials and include families 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[2] have tried to accelerate this search by creating large databases of quantum simulations for both existing and hypothetical materials. These databases give us direct access to the crystal and electronic structure of the materials. Converting the electronic structure to experimental performance, however, typically requires additional work. Take for instance thermoelectricity. Thermoelectric performance is measured by the so-called figure of merit zT, which requires not only the static band structure of a material, but information on both electronic and thermal conductivity. These properties have been calculated for thousands of materials[3], but each additional material requires intensive calculations. Quickly screening many hypothetical materials thus remains difficult.

What if we could skip all that? Accurately determining properties like conductivities requires complex calculations, but somehow the conductivity must result from the electronic structure, which itself must be directly related to the positions of the atoms in the material. This means that functions must exist that map the positions and types of the atoms, directly to the band structure or even the properties. This function is, however, far too complex to be devised by humans. Luckily, we now have a new partner in our quest for new materials: deep learning. Deep learning is a machine learning technique specializing in extracting hidden patterns from large datasets and storing them in complex non-linear functions. Given enough training data, deep learning has shown similar performance to quantum mechanical methods; with a speed gain of several orders of magnitude[4,5]. This makes it possible to finally create quick, yet accurate, structure-property models which we can be directly applied to existing databases to find new energy materials.

Goal
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. The initial goal will be on thermoelectrics, for which machine learning models have already shown success in industry. [6]

Key points:
· Develop and optimize deep learning models that link the crystal structure directly to energy-related properties such as the bandgap, thermal and electronic conductivity and the figure of merit zT.

· Apply the developed models to new structural databases, including our in-house database of new quaternary materials, predicted to be stable.

· 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.

Aspects

Master of Science in Engineering Physics:This thesis subject is closely related to the following clusters of elective courses: MODELLING, MATERIALS and NANO. Physics aspects: based on insights into the quantum physics of solids and their electronic structure; Engineering aspects: application to the properties of materials, use of machine learning

Master of Science in Sustainable Materials Engineering: Establishing and using structure-property relations for materials discovery and materials optimization