Title: Machine Learning Revolutionizes Materials Modeling with Accurate Electronic Structure Calculations
Researchers at the Helmholtz-Zentrum Dresden-Rossendorf (HZDR) in Germany and Sandia National Laboratories in the United States have introduced a groundbreaking machine learning-based simulation technique that outperforms traditional methods for electronic structure calculations. This development opens up new possibilities for advancing fundamental and applied research in various fields, including drug design and energy storage.
The electronic structure, which refers to the arrangement of electrons in matter, plays a significant role in understanding the behavior and properties of materials. However, the lack of a simulation technique that combines high accuracy and scalability across different time and length scales has hindered progress in this field. To overcome this hurdle, the researchers developed the Materials Learning Algorithms (MALA) software stack, which enables access to previously unattainable length scales.
Traditionally, simulating electronic structures has been challenging due to the limitations of existing techniques. Classical atomistic simulation methods can handle complex systems but do not account for quantum electronic structure, while first principles methods, which provide high fidelity, are computationally demanding. The widely used density functional theory (DFT) method, for example, faces restrictions in its predictive capabilities due to its computational limitations with larger systems.
The MALA software stack offers a hybrid approach that combines machine learning with physics-based algorithms. By inputting the arrangement of atoms in space, MALA generates fingerprints called bispectrum components, which encode the spatial arrangement of atoms around a Cartesian grid point. The machine learning model in MALA is trained to predict the electronic structure based on this atomic neighborhood. Crucially, MALA’s machine learning model is independent of system size, allowing it to be trained on data from small systems and deployed at any scale.
The researchers demonstrated the efficacy of MALA by achieving a remarkable speedup of over 1,000 times compared to conventional algorithms for smaller system sizes containing up to a few thousand atoms. Additionally, MALA proved capable of accurately performing electronic structure calculations for larger systems involving over 100,000 atoms, with modest computational effort. This achievement highlights the limitations of conventional DFT codes and showcases the transformative potential of MALA.
The developers anticipate that MALA will significantly enhance electronic structure calculations, enabling researchers to tackle a broader range of societal challenges. These include developing new vaccines and energy storage materials, simulating semiconductor devices on a large scale, studying material defects, and exploring chemical reactions for mitigating carbon dioxide emissions.
Moreover, MALA’s approach aligns well with high-performance computing (HPC), leveraging distributed accelerators to process independent grid points in parallel. As the system size increases, MALA’s ability to harness HPC resources ensures unparalleled speed and efficiency in electronic structure calculations.
Alongside HZDR and Sandia National Laboratories, institutions and companies such as the Georgia Institute of Technology, the North Carolina A&T State University, Sambanova Systems Inc., and Nvidia Corp. are already utilizing MALA for their research and development endeavors.
This significant breakthrough in materials modeling and electronic structure calculations has the potential to revolutionize numerous scientific and technological fields. By combining machine learning with physics-based algorithms, researchers can now delve into unprecedented computational possibilities and address complex challenges with greater precision and efficiency.
For more information, please contact:
Dr. Attila Cangi
Acting Department Head
Center for Advanced Systems Understanding (CASUS) at HZDR
Email: a.cangi@hzdr.de
Research paper:
L. Fiedler, N. A. Modine, S. Schmerler, D. J. Vogel, G. A. Popoola, A. P. Thompson, S. Rajamanickam, A. Cangi, Predicting electronic structures at any length scale with machine learning, npj Computational Materials, 2023
http://doi.org/10.1038/s41524-023-01070-z
