Machine Learning of Exchange Energy Density and Integration with Python-based Electronic Structure Codes

  • Name: Machine Learning of Exchange Energy Density and Integration with Python-based Electronic Structure Codes

Overview of the project

Machine learning is increasingly used to extend the reach of quantum-chemistry methods for molecules and materials. However, applying ML to overcome the limitations of human-designed density functional approximations (DFAs) remains challenging because both DFAs and ML-DFAs still show limited transferability to unseen systems. In this project, with Epicure support, a Python-based framework was developed for learning exchange energy densities and integrating them with electronic-structure codes. Within this framework, we implemented local-energy training (LES) and demonstrated, in subsequent work published in Nature Communications, that real-space machine learning can significantly improve the transferability of density functionals and strengthen ML approaches in quantum chemistry.

 

How did EPICURE support the project and what were the benefits of the support?

“With support from EPICURE, our team developed a Python-based framework that teaches machine-learning models to understand the local energy landscape inside molecules. Instead of learning only final energies, our method learns how the energy is built up point by point in space, using a physically meaningful gauge that assigns energy contributions consistently across space. This richer form of learning greatly improves a model’s ability to predict chemical behaviour it has never seen before, and formed the foundation of results published in Nature Communications.

EPICURE’s contribution was crucial for achieving this. Their experts helped improve the performance and memory efficiency of our implementation, allowing the framework to generate the data volumes required for real-space energy learning, which in turn resulted in training models more smoothly and reliably. These optimisations enabled us to explore more complex ML strategies, test a wider range of architectures, and carry out the computations needed to validate our real-space approach.

This collaboration shows how targeted HPC support can directly accelerate scientific discovery. By improving the scalability and efficiency of our methods, EPICURE made it possible to advance more accurate and more transferable quantum-chemistry tools, opening new possibilities for research in catalysis, materials design and molecular science.”

 

Additional references

Polak, E., Zhao, H. & Vuckovic, S. Real-space machine learning of correlation density functionals. Nature Communication (2025). https://doi.org/10.1038/s41467-025-66450-z

Contact the project:

  • Stefan Vučković (stefan.vuckovic@unifr.ch)
  • Elias Polak (elias.polak@unifr.ch)