Atom-Based Machine Learning for Decomposed Quantum Chemistry
Quantum chemical simulations are a valuable tool, allowing us to explore a vast number of chemicals, with applications within drug discovery and material design. While quantum chemical simulations are generally faster than traditional laboratory experiments, accurate methods are limited to certain molecular sizes due to the computational cost. Machine learning (ML) can alleviate this by training on quantum chemical data and providing predictions on new molecules at a fraction of the computational cost.
In my thesis, we use atom-based neural networks to directly predict quantum chemical properties. We utilize decomposed quantum chemical data, giving us atom-level data instead of molecular-level data. We also continue the development of workflows where ML is used to augment quantum chemical calculations, instead of bypassing them. This maintains the physics-based nature of the quantum chemical method, but still provides computational speedups.
We find our new neural networks provide more physically meaningful results with atomic-level insights into the ML predictions. We further find improvements in the generalizability of our ML models to unseen molecules. Lastly, we can enhance a common quantum chemical workflow by predicting the best method for a given molecule and property.
Principal Supervisor:
Associate Professor Janus J. Eriksen, DTU Chemistry
Co-supervisor:
Professor Sonia Coriani, DTU Chemistry
Examiners:
Professor Thomas Bligaard, DTU Energy
Group leader Matthias Rupp, Luxembourg Institute of Science &Technology, Luxembourg
Professor Jan H. Jensen, University of Copenhagen
Chairperson:
Associate Professor Niels Engholm Henriksen, DTU Chemistry