About two years ago, scientists at CASUS teamed up with collaborators at Sandia National Laboratories (SNL) in the US to initiate the development of a machine learning (ML) framework for building efficient surrogate models of quantum mechanical simulations. Dubbed MALA (short for MAterials Learning Algorithms), the first version of the framework was made available to the public in the fall of 2021. After releasing an updated version one year later, the MALA team is now hosting the inaugural MALA hackathon to enhance and extend their software stack.
The first MALA hackathon was a two-day event in early January 2023 designed to improve the efficiency and capabilities of the MALA framework, a cutting-edge scientific software package for generating machine learning (ML) surrogate models. These models are used to replace time-consuming and resource-intensive density functional theory calculations, allowing for faster and more accurate simulations of materials properties and their electronic structure across a wide range of pressures and temperatures.
The MALA framework is built as a modular and open-source python package, enabling researchers to perform the entire modeling toolchain with minimal coding. This makes it particularly useful for large-scale simulations that go beyond what is achievable with state-of-the-art codes by leveraging physics-informed machine learning. “The first results of the MALA project are very promising,” says Lenz Fiedler, a Ph.D. student at CASUS and a MALA core developer. “With this hackathon, we aim to take the next step in achieving an efficient end-to-end machine learning workflow, and we’re excited to see the results,” Lenz adds.
The ten participants of the hackathon focused on utilizing the full capabilities of graphical processing units (GPU) for inference, streamlining data loading, and expanding the front-end capabilities of the software. Interestingly, MALA also has raised the interest of US-based GPU designer and artificial intelligence specialist Nvidia Corp.: Some of their developers actively work on accelerating MALA on Nvidia’s GPU hardware.
The results of the hackathon are expected to be incorporated into an upcoming release of the MALA package, to be published as early as February 2023.