Thermoelectric material screening with combined machine learning and atomistic modelling

Thermoelectric material screening with combined machine learning and atomistic modelling

Thermoelectric module

Thermoelectric materials can turn heat into electricity, but better materials are needed for large-scale recovery of waste heat.

Our project aims to identify materials with great promise using only computer experiments and machine learning.

prosjekt

About/Aims
Background
Thermoelectric modules have the ability to turn heat into electricity and vice versa. This ability makes them useful in many niche technologies. Examples include power sources for Mars rowers, wine-coolers, and smart-phone chargers for campers. 
 
The potential of thermoelectrics goes far beyond niche applications. Consider the fact that about 70% of all energy generated in various industrial processes and transportation is lost as waste heat. If we could recover even a small fraction of this waste heat, it would open one more front in the fight against the rising global greenhouse gas emissions. But first, thermoelectrics need higher efficiency and this demands developing better thermoelectric materials. Such material development has unfortunately historically been very slow. Part of the reason is that material development is always costly and arduous. But, this is particularly so for thermoelectric materials, due to their very specific set of properties. They must be poor thermal conductors and good electrical conductors. At the same time, heat gradients should induce highly directional electrical flow. Either, the electrical charge should flow from the hot end to the cold of a thermoelectric material, or the other way around. These requirements make it hard to use simple intuitive arguments to improve materials. Luckily, methods such as density functional theory (DFT) can be used to compute material properties without the need for any experimental input. Thus, far more materials can be studied than in experimental studies alone. With the adoption of such an approach, the number of new promising thermoelectrics materials have been rising faster than ever.
 
While DFT methods greatly expand our ability to predict material properties, they are limited by considerable computational costs.  Ideally, we would like to explore the properties of all  100,000 known inorganic compounds and identify the very best ones --- not to mention exploring the far larger number of proposed hypothetical materials yet be synthesized. This is simply too costly to model. 
 
In this project, we will attempt to overcome limits set by DFT and adopt machine learning methods trained on DFT data. The goal is to develop algorithms that can pinpoint materials worthy of further exploration. Based on this, we aim to recommend new materials with great potential so they can be synthesized by experimentalists and hopefully be adopted in thermoelectric modules in the future. 
Objectives

-Predict new promising thermoelectric materials.

- Develop effective machine-learning schemes to estimate thermoelectric properties. 

More about the project

The project involves a collaboration between the Material Theory group and the data science group at NMBU and with Ole Martin Løvvik at SINTEF