How selvachandrasekaranselvaraj/deepatom works
Project type and purpose
• DeepAtom is a standalone Python library/package for materials-science machine-learning workflows: loading trajectory data, computing atomic descriptors, preparing training data, training/predicting with a graph neural network (GNN).
• It is not a web/full-stack app or microservice: it’s a modular scientific pipeline accessed via a CLI or imported as a library.
File structure
• Top-level files: setup.py, requirements.txt, README.md define packaging and dependencies.
• Core package directory (deepatom/):
• main.py: orchestrator and CLI entry point (DeepAtom class).
• descriptors.py: descriptor calculators (Coulomb, ACSF, SOAP, MBTR, etc.).
• gnn.py: graph-neural-network architecture, training, saving/loading.
• utils.py: data-handling utilities for parsing LAMMPS trajectory files.
• tools/: shell and Perl scripts for pre/post-processing outside Python.
• tests/: unit tests for descriptors (and potentially other modules).
README for additional insights
• Workflow stages clearly defined: load data → calculate descriptors → prepare data → train GNN → predict.
• External input: LAMMPS trajectory files.
• Outputs: descriptor arrays, saved model file (deepatom_model.pth), prediction metrics.
• Dependencies: NumPy, PyTorch (for GNN), scikit-learn (likely for scaling/splitting), perhaps ASE for structure parsing.
Detail
• “Tests” component connected to the modules it validates.
• Packaging/deployment layer (setup.py, pip install) to show how the library is distributed.