Machine learning takes materials modeling into new era

Newswise — The electronic structure, which refers to the arrangement of electrons in matter, plays a critical role in both fundamental and applied research areas like drug design and energy storage. However, the progress in these technologies has been hindered by the lack of a simulation technique that offers both high fidelity and scalability across various time and length scales. Recently, a team of researchers from the Center for Advanced Systems Understanding (CASUS) at the Helmholtz-Zentrum Dresden-Rossendorf (HZDR) in Görlitz, Germany, and Sandia National Laboratories in Albuquerque, New Mexico, USA, has introduced a groundbreaking machine learning-based simulation method that surpasses traditional electronic structure simulation techniques. By developing the Materials Learning Algorithms (MALA) software stack, they have unlocked access to previously unattainable length scales.

Electrons, as elementary particles, hold significant importance. Their quantum mechanical interactions with each other and atomic nuclei give rise to a wide range of phenomena observed in the fields of chemistry and materials science. Understanding and controlling the electronic structure of matter offer insights into various aspects such as molecular reactivity, energy transport within planets, and the mechanisms of material failure.

In order to tackle scientific challenges, computational modeling and simulation have increasingly become essential, leveraging the capabilities of high-performance computing. However, achieving realistic simulations with quantum precision faces a significant hurdle: the lack of a predictive modeling technique that combines high accuracy with scalability across different time and length scales. Classical atomistic simulation methods can handle large and complex systems, but their exclusion of quantum electronic structure limits their applicability. On the other hand, simulation methods based on first principles, which do not rely on assumptions like empirical modeling and parameter fitting, provide high fidelity but are computationally intensive. For example, density functional theory (DFT), a widely used first principles method, exhibits cubic scaling with system size, which restricts its predictive capabilities to small scales.

Hybrid approach based on deep learning

The research team has recently introduced an innovative simulation method called the Materials Learning Algorithms (MALA) software stack. In the field of computer science, a software stack refers to a collection of algorithms and software components combined to create a software application that addresses a specific problem. Lenz Fiedler, a Ph.D. student and key developer of MALA at CASUS, explains that MALA integrates machine learning with physics-based approaches to predict the electronic structure of materials. It employs a hybrid approach, utilizing deep learning, an established machine learning method, to accurately predict local quantities, while employing physics algorithms to compute global quantities of interest.

The MALA software stack takes the spatial arrangement of atoms as input and generates fingerprints called bispectrum components. These components encode the atomic arrangement around each point on a Cartesian grid. The machine learning model in MALA is trained to predict the electronic structure based on these atomic neighborhoods. An important advantage of MALA is that its machine learning model is independent of the system size, allowing it to be trained on data from small systems and applied at any scale.

In their publication, the research team demonstrated the impressive effectiveness of this approach. They achieved a speedup of over 1,000 times for smaller system sizes, consisting of a few thousand atoms, compared to conventional algorithms. Furthermore, the team showcased MALA's ability to accurately perform electronic structure calculations at a large scale, involving over 100,000 atoms. Notably, this achievement was accomplished with modest computational effort, highlighting the limitations of traditional density functional theory (DFT) codes.

Attila Cangi, the Acting Department Head of Matter under Extreme Conditions at CASUS, explains that as the system size increases and more atoms are involved, DFT calculations become impractical, while MALA's speed advantage continues to grow. The key breakthrough of MALA lies in its ability to operate on local atomic environments, enabling accurate numerical predictions that are minimally influenced by system size. This groundbreaking achievement opens up computational possibilities that were once considered unattainable.

Boost for applied research expected

Cangi is determined to push the boundaries of electronic structure calculations by harnessing the power of machine learning. He anticipates that MALA will ignite a transformative shift in electronic structure calculations, as it enables simulations of significantly larger systems at an unprecedented speed. This breakthrough opens up new possibilities for addressing various societal challenges, including the development of vaccines and novel materials for energy storage, conducting large-scale simulations of semiconductor devices, studying material defects, and exploring chemical reactions to convert carbon dioxide into environmentally friendly minerals.

Additionally, MALA's approach is highly compatible with high-performance computing (HPC). As the system size expands, MALA facilitates independent processing on the computational grid it utilizes, effectively harnessing HPC resources, particularly graphical processing units. Siva Rajamanickam, a staff scientist and parallel computing expert at the Sandia National Laboratories, explains that MALA's algorithm for electronic structure calculations aligns well with modern HPC systems equipped with distributed accelerators. The ability to decompose work and execute parallel computations across different grid points on multiple accelerators makes MALA an ideal choice for scalable machine learning on HPC resources, resulting in unparalleled speed and efficiency in electronic structure calculations.

In addition to the collaborative efforts of HZDR and Sandia National Laboratories, MALA is already being utilized by various institutions and companies, including the Georgia Institute of Technology, the North Carolina A&T State University, Sambanova Systems Inc., and Nvidia Corp.