AI Technology: Deep Learning
Machine learning plays a pivotal role in extending the reach of quantum-chemistry methods for simulating molecules and materials. However, using machine learning to overcome the limitations of human-designed density functional approximations (DFAs), the primary workhorses of quantum simulations, remains challenging because current approaches still show limited transferability to unseen chemical systems.
This project addresses this challenge by developing thermaMLS2, a real-space machine-learned correlation model that builds on our recent work on local-energy learning and perturbation-theory-based density functionals.
The method will combine local-energy training with new descriptors based on thermally assisted occupations, enabling the model to adjust to different correlation regimes. Transition-metal datapoints will also be included directly in the training, extending applicability beyond the main-group chemistry on which most ML functionals are based.
The resulting model is expected to retain the computational cost of double-hybrid DFAs while offering improved reliability on challenging systems. Large-scale GPU resources provided by EuroHPC are essential for exploring the relevant DFA design space enabling the efficient training and validation of thermaMLS2.
Stefan Vuckovic, University of Fribourg, Switzerland