Machine learning for computational materials science and chemistry with MALA. A Basic Fully Online Pipeline and Framework
In this project we show how a Materials Learning Algorithms (MALA) Framework and Pipeline works. From the generation of the DFT data with Quantum Espresso, the conversion via descriptors to MALA, the training and testing of a neural network and finally the hyperparameter optimisation of the neural network (TODO in v3).
We show a framework that can run locally on an Ubuntu machine, on a Windows machine using Google Colaboratory's free cloud computing framework for Quantum Espresso and a locally installed MALA, and finally in a fully online pipeline using Google's resources for both DFT simulations and neural network training.
DFT simulations are created using Google Colab's tool, we focus on simulating both Be2 and Si2 systems's LDOS Colab File. The files generated will be randomized by varying randomly Cell Parameters or Atomic Positions. Files are then downloaded in Zip for following processing. Only needs the Pseudopotencial files included Si pseudo Be pseudo
MALA simulations are able to be run both through COLAB's environment in Colab File and locally with the notebook. For running locally, only a MALA installation and Pytorch are needed, and works both in Windows and Linux Environments. LAMMPS is optional but not needed as it needs special compiling instructions.
For Data conversion there is both a tool in the notebook and (as it is a slow process) optional .py files befile sifile to copy into materials folder to generate MALA conversion files using Multiprocessing. It is needed to create the /snapshots folder inside the materials folder manually.
If you only want to try the MALA ML models, there are also files uploaded for hands-on work.
We use a NN in Mala to predict the LDOS of the different snapshots as a proof of concept of the MALA environment and framework. NN are very simple but are able to predict within <10meV/at the energy bands in Be2 systems and within <60meV in Si2 systems. Showing a clear improvement over other ML-DFT algorithms