Machine learning-based simulation method for predicting electronic structure; created by the Center for Advanced Systems Understanding and Sandia National Laboratories’ MALA project. By combining machine learning with physics algorithms, MALA achieves superior performance compared to conventional approaches, with simulation speeds of over 1,000 times for tiny systems and the ability to accurately model large-scale systems of over 100,000 atoms. This breakthrough is optimized for use with high-performance computers and will change the face of applied research forever.
Deep learning approach enables accurate electronic structure calculations at large scales.
Understanding the electrical structure of a substance is vital not only for basic science but also for practical applications like medication design and energy storage. However, the development of such systems has been hampered by the absence of a simulation technique that provides both high fidelity and scalability across many time and length scales. Scientists at Sandia National Laboratories in Albuquerque, New Mexico, and the Center for Advanced Systems Understanding (CASUS) at the Helmholtz-Zentrum Dresden-Rossendorf (HZDR) in Görlitz, Germany, have developed a machine learning-based simulation method (npj Computational Materials, DOI: 10.1038/s41524-023-01070-z) that outperforms conventional electronic structure simulation methods. Using their Materials Learning Algorithms (MALA) software stack, researchers can gain access to length scales that were previously inaccessible.
Electrons are fundamentally important elementary particles. They cause a wide variety of phenomena in chemistry and materials science through their quantum mechanical interactions with one another and with atomic nuclei. Insights into molecular reactivity, planetary structure and energy flow, and material failure mechanisms can be gained by studying and manipulating the electrical structure of matter.
Computational modeling and simulation, made possible by high-performance computing, are becoming increasingly popular methods for tackling complex scientific problems. The lack of a predictive modeling approach that combines high accuracy with scalability across diverse length and time scales is a major roadblock on the path to producing realistic simulations with quantum precision. While classical atomistic simulation methods are capable of simulating complex systems, they are limited in their utility because they ignore the quantum electronic structure of the system. In contrast, first principles methods of simulation, such empirical modeling and parameter fitting, yield excellent fidelity but are computationally intensive. Density functional theory (DFT), a popular first principles approach, suffers from the same limitation because of its cubic scaling with system size.
More than ten thousand virtual beryllium atoms, captured in an instantaneous deep learning simulation. This material’s electron distribution is shown as a series of red (representing delocalized electrons) and blue (representing electrons near to the atomic nuclei) point clouds. Traditional DFT calculation is infeasible for this simulation. The task was completed in roughly 5 minutes using only 150 CPUs thanks to MALA. To make the simulation more understandable, graphical filters have been applied. The filters are also to blame for the white edges. The underlying structure provides clues about how deep learning operates. HZDR / CASUS Gets the Credit
Hybrid approach based on deep learning
A new simulation approach, the Materials Learning Algorithms (MALA) software stack, has been introduced by the study team. A software stack, in the field of computer science, refers to the collection of methods and software components used to build a program that addresses a certain problem.
To forecast the electrical structure of materials, MALA mixes machine learning with physics-based techniques, as CASUS Ph.D. student and key developer Lenz Fiedler explains. It uses a combination of physics methods for computing global quantities of interest and a well-established machine learning method called deep learning for precise local quantity prediction.
With the atomic arrangement as input, the MALA software stack creates fingerprints called bispectrum components that encode the atomic arrangement around a Cartesian grid point. In MALA, the atomic neighborhood is used as training data for a machine learning model that predicts the electronic structure. MALA’s machine learning model is a major strength since it can be trained using data from low-scale systems and then used in large-scale deployments.
The study group demonstrated the strategy’s outstanding efficacy in a published paper. For systems with up to a few thousand atoms, they were able to achieve a speedup of over a thousand times when compared to traditional techniques. The scientists also showed that MALA can accurately execute electronic structure computations for systems with more than 100,000 atoms, a feat that would have been impossible with previous methods. The success with relatively little computing effort is noteworthy because it exposes the shortcomings of traditional DFT codes.
Acting CASUS Matter under Extreme Conditions Department Head Attila Cangi says, “As the system size increases and more atoms are involved, DFT calculations become impractical, whereas MALA’s speed advantage continues to grow.” The capacity of MALA to operate on local atomic environments is its greatest innovation; this allows for precise numerical forecasts that are only little affected by the scale of the system. New computational frontiers are now within reach, thanks to this historic achievement.
Boost for applied research expected
Cangi plans to use machine learning to advance electronic structure calculations, saying, “We anticipate that MALA will spark a transformation in electronic structure calculations, as we now have a method to simulate significantly larger systems at an unprecedented speed.” Researchers will have a much better starting point in the future, allowing them to tackle a wide variety of societal challenges such as creating new vaccines and novel materials for energy storage, running large-scale simulations of semiconductor devices, investigating material defects, and investigating chemical reactions for converting atmospheric greenhouse gas carbon dioxide into climate-friendly minerals.
In addition, the methodology employed by MALA works exceptionally well with HPC. MALA allows for decoupled processing on the computational grid it employs, making efficient use of high-performance computing (HPC) resources, in particular graphics processing units, as the system’s size increases.
“MALA’s algorithm for electronic structure calculations maps well to modern HPC systems with distributed accelerators,” says Siva Rajamanickam, a staff scientist and expert in parallel computing at Sandia National Laboratories. MALA is a perfect fit for scalable machine learning on HPC resources, resulting in unprecedented speed and efficiency in electronic structure calculations, thanks to its ability to decompose work and execute in parallel different grid points across different accelerators.