Machine Learning for Materials Prediction


In the process of building surrogate machine-learning models for the prediction of materials properties, we have previously contributed an iteration on the well-established radial distribution function.
Currently we are researching how we can improve prediction accuracy on various key properties (bandgap and stability) while keeping the amount of (expensive, ab-initio-simulated) training data low.
We focus on autoencoder-neural-networks to achieve this goal, aiming to compress the "raw" fingerprint into a meaningful intermediary and ideally opening up the way to generate new materials.
Covered Topics: Density Functional Theory, Machine Learning, Data science, Molecular Dynamics
Contact Person: Felix Mayr, Alessio Gagliardi