atomistic-machine-learning / schnetpack

SchNetPack - Deep Neural Networks for Atomistic Systems
Other
751 stars 210 forks source link

Added steps for user controlled data screening and hyper parameter testing #615

Closed santos-adh114 closed 3 months ago

santos-adh114 commented 3 months ago

There are few updates on the branch "santosh". While some changes are trivial (changing path of config file, dataset etc.,). There are major changes in the following files compared to the main branch.

  1. schnetpack/examples/wannier/wannier_centers_generator.py Here the function "read_data()" in line 241 is updated. The update scans all 1593 neutral configurations (wannier centers with no external electric field) made publicly available by Remsing et al (https://zenodo.org/records/5760191) to retain only selected configurations for model training, validation and testing. The selection is such that (a) each oxygen atoms in the given configurations has four wannier centers and (b) each of those four wannier centers are within 0.74 Angstrom (tunable parameter) from the associated oxygen atom . For reference, each of the configurations in this dataset has 64 oxygen atoms, 128 hydrogen atoms and 256 (ideally 4 for each oxygen) wannier centers.

  2. schnetpack/examples/wannier/submit_jobs_sabine.py This is the modified job submission file (specifically designed to work in carya/sabine clusters). As earlier it uses the package minilaunch to control the job submission. This script has four loops corresponding to cutoff radius (cutoff :[4.0,5.0,6.0], number of layers (n_interactions): [5,6,7,8,9,10], number of basis to represent atoms (n_atom_basis: [16,24,32,40,48] and number of gaussian functions for modeling radial distribution of the given atom (n_brf: [16,24,32,40,48]. Each jobs are submitted in the directory schnetpack/examples/wannier/jobdir-tmp/. Each subfolders within this directory are named as "rcut{r_cut}nlay{n_interactions}_atmbas{n_atom_basis}_ngaus{n_rbf}"