yqshao/deltaml repository overview

⁠Delta ML with PiNN

⁠Introduction

To be written.

⁠Run this project

Before running anything, download the dataset⁠ and put it into the dataset folder (name as "revPBE0-D3.data").

url="https://archive.materialscloud.org/record/file?filename=training-set.zip&file_id=4887cbec-a3b0-48f0-9e4b-1836c84a85b8&record_id=71";
curl -sSL $url -o training-set.zip && unzip training-set.zip &&\
mkdir -p datasets && mv training-set/input.data datasets/revPBE0-D3.data &&\
rm -r training-set*

The project contains Nextflow⁠ scripts and a Docker image which should make it easy to reproduce, the scripts used in are plain bash/python scripts.

For now the workflow is separated into three stages:

• label.nf: computes DFTB+ labels for the entire dataset
• train.nf: train the Delta-ML model
• mdrun.nf: produces and analyzes the MD trajectory

To run the project locally:

nextflow run label[train,mdrun].nf -resume

To run the project on a HPC cluster (assuming the queuing and project info in nextflow.config are correct)

nextflow run label[train,mdrun].nf -profile rackham -resume

A Docker image containing all the requirements for this project is continuously built for this project, and the Nextflow script will use a singluarity⁠ image built out of it by default.

⁠Folder structure

• nextflow.config: config for running locally or on a HPC cluster
• python/: Delta-ML calculator for ASE

inputs

• inputs: input files including that for PiNN and DFTB

outputs

• datasets: downloaded/generated datasets
• models: trained models
• trajs: MD trajectories

⁠Roadmap

• Implement label.nf and generate the Delta-ML dataset.
• Train the Delta-ML model.
• Implement the calculator for ASE/DeltaML, and run the MD.