Publications Repository - Helmholtz-Zentrum Dresden-Rossendorf

Ab-initio simulations are crucial tools for many scientific applications, from materials science to drug discovery. This is due to powerful simulation techniques such as Density Functional Theory (DFT), that combine high accuracy with computational feasibility. Yet, there exist applications unattainable to even the most performant of DFT programs. A prominent example is the modeling of materials on multiple time and length scales, especially under ambient or extreme conditions. While these simulations hold the potential to both further our understanding of important physical phenomena such as planetary formation or radiation damages in fusion reactor wall, they evade traditional ab-initio approaches due to their size and complexity.
Surrogate models can mitigate these computational restrictions, by reproducing DFT-level results at a fraction of the cost. Here, were present the Materials Learning Algorithms (MALA) package, an open source python package for building neural network based surrogate models for materials science. MALA provides easy-to-use functions to process DFT data, build models and use these models to replace DFT calculations, as shown for simulations of Aluminium at both 298K and 933K, as well as Iron at 3000K. The source code for MALA is publicly available on Github and developed by the Center for Advanced Systems Understanding (CASUS), Sandia National Laboratories, and Oak Ridge National Laboratory.