Machine learning incorporates material modeling

The arrangement of electrons in materials, known as electronic structure, plays an important role not only in basic research but also in applied research such as drug discovery and energy storage. However, the lack of simulation techniques that provide both high fidelity and scalability over various time and length scales has long hindered progress in these techniques. Researchers at the Center for Advanced Systems Understanding (CASUS) at the Dresden-Rossendorf-Helmholtzzentrum (HZDR) in Görlitz, Germany, and Sandia National Laboratories, Albuquerque, New Mexico, USA, developed a machine-learning-based simulation method. developed (npj calculation material, DOI: 10.1038/s41524-023-01070-z) replaces traditional electronic structure simulation techniques. The company’s Materials Learning Algorithm (MALA) software stack enables access to previously unattainable length scales.

Electrons are fundamentally important elementary particles. Their interactions and quantum mechanical interactions with atomic nuclei give rise to numerous phenomena observed in chemistry and materials science. Understanding and controlling the electronic structure of matter provides insight into molecular reactivity, intraplanetary structure and energy transport, and matter breakage mechanisms.

Scientific challenges are increasingly being addressed through computational modeling and simulation, leveraging the power of high-performance computing. However, a major obstacle to achieving realistic simulations with quantum precision is the lack of predictive modeling techniques that combine high accuracy with scalability over various length and time scales. Classical atomic simulation methods can handle large and complex systems, but the omission of the quantum electronic structure limits their applicability. Conversely, assumption-independent simulation methods (first-principles methods), such as empirical modeling and parameter fitting, offer high fidelity but are computationally intensive. For example, the widely used first-principles method, density functional theory (DFT), exhibits his cubic scaling with system size, thus limiting its predictive power to small scales.

Hybrid approach based on deep learning

The research team now presents a new simulation technique called the Material Learning Algorithm (MALA) software stack. In computer science, a software stack is a collection of algorithms and software components that combine to create a software application to solve a specific problem. “MALA integrates machine learning and physics-based approaches to predict the electronic structure of materials. It employs a hybrid approach that utilizes well-established machine-learning techniques, called , to accurately predict local quantities, complemented by physical algorithms to compute global quantities of interest.”

The MALA software stack takes as input the arrangement of atoms in space and produces fingerprints known as bispectral components. It encodes the spatial arrangement of atoms around a Cartesian lattice point. MALA’s machine learning model is trained to predict the electronic structure based on this atomic neighborhood. A major advantage of MALA is that machine learning models are system-scale agnostic, meaning they can be trained on data from small systems and deployed at any scale.

A team of researchers demonstrated the remarkable effectiveness of this strategy in a publication. They achieved speedups of more than 1,000 times for small system sizes consisting of up to thousands of atoms compared to conventional algorithms. Additionally, the team demonstrated MALA’s ability to accurately perform large-scale electronic structure calculations involving more than 100,000 atoms. Remarkably, this result was achieved with a modest amount of computation, revealing the limitations of his conventional DFT code.

Attila Cangi, Deputy Head of Extreme Conditions Materials Division at CASUS, explains: “DFT calculations become impractical as the size of the system increases and more atoms are involved.Meanwhile, the speed advantage of MALA continues to grow. “The ability to operate in a local atomic environment allows for accurate numerical predictions that are minimally affected by the scale of the system. This breakthrough was previously unattainable.” It opens up computational possibilities that were thought to be

Promotion of applied research is expected

Cangi aims to leverage machine learning to push the boundaries of electronic structure calculations. “We expect MALA to revolutionize electronic structure calculations, as we now have a way to simulate very large systems at unprecedented speed. Based on this baseline, we can address a wide range of societal challenges, such as developing new vaccines and new materials for energy storage, conducting large-scale simulations of semiconductor devices, studying material defects, and exploring chemical reactions to transform the atmosphere. We will be able to deal with it, turning carbon dioxide, a greenhouse gas, into a climate-friendly mineral.”

Moreover, MALA’s approach is particularly well-suited for high-performance computing (HPC). As systems grow in size, MALA enables independent processing on the available computational grid, effectively utilizing HPC resources, especially graphics processing units. Shiva Rajamanikam, staff scientist and parallel computing expert at Sandia National Laboratories, said, “MALA’s electronic structure computation algorithms are well suited for modern HPC systems with distributed accelerators. and the ability to run in parallel across different grid points.” Accelerators make MALA ideally suited for scalable machine learning on HPC resources, delivering unrivaled speed and efficiency in electronic structure computations. . ”

Apart from development partners HZDR and Sandia National Laboratories, MALA has already been adopted by institutions and companies such as Georgia Tech, North Carolina A&T State University, Sambanova Systems Inc. and Nvidia Corp.