We develop the Materials Learning Algorithms (MALA), a physics-informed machine learning framework that aims to accelerate conventional density functional theory simulations. Using neural networks, MALA efficiently computes the electronic structure of matter, enabling accurate determination of energies and forces that are critical for atomistic simulations.
MALA is a scalable method that balances accuracy and speed, overcoming the limitations of conventional density functional theory simulations. It combines the scalability of atomistic simulations with the high accuracy of first-principles methods, paving the way for electronic structure calculations at unprecedented length and time scales. This advancement in materials modeling opens up a broad range of potential applications.
Dr Attila Cangi
Bartosz Brzoza
Karan Shah
Dr Lenz Fiedler
Dr Kushal Ramakrishna
The arrangement of electrons in matter, known as electronic structure, plays a critical role in fundamental and applied research such as drug design and energy storage. However, the lack of a simulation technique that provides both high fidelity and scalability across different time and length scales has long been an obstacle to the progress of these technologies. We have developed a machine learning method – the Materials Learning Algorithms (MALA) – that replaces traditional electronic structure simulation techniques. MALA enables electronic structure simulations at previously unattainable length scales.
Predicting the Electronic Structure of a Stacking Fault
The MALA machine learning model enables us to make accurate predictions of the electronic structure beyond the length scales feasible with standard electronic structure methods like density functional theory. By leveraging the principle of nearsightedness of the electronic structure, MALA is trained on local electronic environments to make predictions on much larger length scales. We demonstrate this capability by predicting the electronic structure across a stacking fault in a slab of beryllium.
Predicting the electronic structure of a stacking fault in a slab of beryllium. The simulation contains 131,072 beryllium atoms, with a stacking fault localized in the center of the slab, where the local atomic arrangement exhibits fcc geometry, deviating from the bulk hcp symmetry.
Publications
Machine learning the electronic structure of matter across temperatures
Predicting electronic structures at any length scale with machine learning
Training-free hyperparameter optimization of neural networks for electronic structures in matter
Accelerating finite-temperature Kohn-Sham density functional theory with deep neural networks
CASUS
Untermarkt 20, D-02826 Görlitz
Conrad-Schiedt-Str. 20, D-02826 Görlitz
About
Events & Seminars
Scultetus
Publications
Career
Imprint
Privacy Policy
CASUS is an institute of Helmholtz-Zentrum Dresden-Rossendorf (HZDR)
Funded by:
in close cooperation with:
Funded by:
In close cooperation with: