MALA is a scalable machine-learning framework that reproduces key density-functional-theory outputs for large atomistic systems, from local densities of states to total energies.
MALA learns the mapping from local atomic environments to electronic observables, so a simulation can obtain densities of states, electron densities and total energies without solving the Kohn–Sham problem at every step. The package covers the whole workflow — sampling training data, fitting models and running inference in parallel — and is built to interoperate with standard DFT and molecular-dynamics codes rather than replace them.
The paper demonstrates the framework on boron clusters, on aluminum across its solid–liquid phase boundary, and on the electronic structure of a stacking fault in a large beryllium slab. Scaling tests document the speedup achieved and expose the remaining bottlenecks, giving a realistic picture of when learned electronic structure becomes the cheaper option. The overall aim is to push electronic-structure accuracy into system sizes that conventional DFT cannot reach.