Data

Tools

Indexes

Applications

Experiments

Open Source Science

tensormol

Tensorflow + Molecules = TensorMol

https://github.com/jparkhill/tensormol

Science Score: 33.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • ○
    CITATION.cff file
  • ○
    codemeta.json file
  • ○
    .zenodo.json file
  • ✓
    DOI references
    Found 6 DOI reference(s) in README
  • ✓
    Academic publication links
    Links to: arxiv.org, rsc.org, acs.org
  • ✓
    Committers with academic emails
    6 of 13 committers (46.2%) from academic institutions
  • ○
    Institutional organization owner
  • ○
    JOSS paper metadata
  • ○
    Scientific vocabulary similarity
    Low similarity (13.5%) to scientific vocabulary

Keywords

chemistry force-field machine-learning meta-dynamics molecular-dynamics molecular-simulation molecules monte-carlo neural-network simulation tensorflow
Last synced: 6 months ago · JSON representation

Repository

Tensorflow + Molecules = TensorMol

Basic Info
Statistics
  • Stars: 275
  • Watchers: 44
  • Forks: 72
  • Open Issues: 18
  • Releases: 0
Topics
chemistry force-field machine-learning meta-dynamics molecular-dynamics molecular-simulation molecules monte-carlo neural-network simulation tensorflow
Created over 9 years ago · Last pushed about 5 years ago
Metadata Files
Readme Contributing License

README.md

►

-Title signature by Alex Graves' handwriting LSTM https://arxiv.org/abs/1308.0850

PyPI version Documentation Status

Authors:

Kun Yao (kyao@nd.edu), John Herr (jherr1@nd.edu), David Toth (dtoth1@nd.edu), Ryker McIntyre(rmcinty3@nd.edu), Nicolas Casetti, John Parkhill (john.parkhill@gmail.com)

Model Chemistries:

  • Behler-Parrinello with electrostatics
  • Many Body Expansion
  • Bonds in Molecules NN
  • Atomwise Forces
  • Inductive Charges

Simulation Types:

  • Optimizations
  • Molecular Dynamics (NVE,NVT Nose-Hoover)
  • Monte Carlo
  • Open/Periodic Boundary Conditions
  • Meta-Dynamics
  • Infrared spectra by propagation
  • Infrared spectra by Harmonic Approximation.
  • Nudged Elastic Band
  • Path integral simulations via interface with I-PI MD engine.

News:

  • Did we disappear? NO! but TensorMol0.2 is being developed in a private developers branch. It'll be back here soon.
  • (update 3/29/2018) TensorMol0.2 will arrive before 5/1. It adds additn'l element support, geometrical constraints, conformational search and other features.

License: GPLv3

By using this software you agree to the terms in COPYING

Installation:

  • Install TensorFlow(>1.1), otherwise TensorMol is self-contained.
  • Works on OSX, Ubuntu, and Windows subsystem for Linux: git clone https://github.com/jparkhill/TensorMol.git cd TensorMol # If you are using python2x sudo pip install -e . # If you are using python3x sudo pip3 install -e . python test.py

Demo of training a neural network force field using TensorMol:

  • Copy the training script into the tensormol folder:cp samples/training_sample.py . Run the script: python training_sample.py This will train a network force field for water.

Test example for TensorMol01:

  • Download our pretrained neural networks (network.tar.gz). Networks for water and molecules that only contains C, H, O, N (The file is about 6 Gigabyte. This may take a while)
  • Copy the zipped trained networks file (network.tar.gz) into TensorMol folder. Unzip it. The networks should be in './networks' folder.
  • Copy the test script into the tensormol folder:cp samples/test_tensormol01.py . Run the script: python test_tensormol01.py The test sample contains geometry optimization, molecular dynamic, harmonic IR spectrum and realtime IR spectrum.

Timing Information

TensorMol is robust and fast. You can get an BP+electrostatic energy and force of this monstrous cube of 24,000 atoms in less than 100 seconds on a 2015 MacbookPro (Core i7 2.5Ghz, 16GB mem). Periodic simulations are about 3x more expensive.

Usage:

  • import TensorMol as tm
  • We are working on /doc/Tutorials, but it's sparse now. We've done a lot of re-writing, and so if you are looking for good examples, look for /samples files with recent commits.
  • A collection of tests are located in samples/test_tensormol01.py but this requires first downloading the trained networks from the Arxiv paper.
  • python samples/test_tensormol01.py
  • IPI interface: start server: ~/i-pi/i-pi samples/i-piinterface/H2Ocluster.xml > log &; run client: python test_ipi.py

Sample Results

Biological molecules

Because Neural network force fields do not rely on any specific atom typing or bond topology, the agony of setting up simulations of biological molecules is greatly reduced. This gif is a periodic optimization of PDB structure 2EVQ, in explicit polarizable TensorMol solvent.

Chemical Reactions

Converged nudged elastic band simulations of the cyclization cascade of endiandric acid C (c.f. K. C. Nicolaou, N. A. Petasis, R. E. Zipkin, 1982, The endiandric acid cascade. Electrocyclizations in organic synthesis. 4. Biomimetic approach to endiandric acids A-G. Total synthesis and thermal studies, J. Am. Chem. Soc. 104(20):5560–5562).

This reaction path can be found in a few minutes on an ordinary laptop. Relaxation from the linearly interpolated guess looks like this:

The associated energy surface is shown below.

Dynamic Properties

  • Tyrosine Harmonic IR spectrum

Publications and Press:

  • John E. Herr, Kun Yao, David W. Toth, Ryker Mcintyre, John Parkhill.Metadynamics for Training Neural Network Model Chemistries: a Competitive Assessment. Arxiv
    • Kun Yao, John E. Herr, David W. Toth, Ryker Mcintyre, John Parkhill. The TensorMol-0.1 Model Chemistry: a Neural Network Augmented with Long-Range Physics. Chemical Science
  • Writeup in Chemistry World
  • Kun Yao, John E. Herr, Seth N. Brown, & John Parkhill. Intrinsic Bond Energies from a Bonds-in-Molecules Neural Network. Journal of Physical Chemistry Letters (2017). DOI: 10.1021/acs.jpclett.7b01072
  • Kun Yao, John Herr, & John Parkhill. The Many-body Expansion Combined with Neural Networks. Journal of Chemical Physics (2016). DOI: 10.1063/1.4973380
  • Kun Yao, John Parkhill. The Kinetic Energy of Hydrocarbons as a Function of Electron Density and Convolutional Neural Networks. Journal of Chemical Theory and Computation (2016). DOI: 10.1021/acs.jctc.5b01011

Requirements:

  • Minimum Pre-Requisites: Python2.7x, TensorFlow sudo pip install tensorflow
  • Also now works in Python3.6.
  • Useful Pre-Requisites: CUDA7.5, PySCF
  • To Train Minimally: ~100GB Disk 20GB memory
  • To Train Realistically: 1TB Disk, GTX1070++
  • To Evaluate: Normal CPU and 10GB Mem

Acknowledgements:

  • Google Inc. (for TensorFlow)
  • NVidia Corp. (hardware)
  • von Lilienfeld Group (for GBD9)
  • Chan Group (for PySCF)

Common Issues:

  • nan during training due to bad checkpoints in /networks (clean.sh)
  • Also crashes when reviving networks from disk.
  • if you have these issues try re-installing or:

sh clean.sh

Owner

  • Name: John Parkhill
  • Login: jparkhill
  • Kind: user
  • Location: Austin, TX

Fmr. Prof. Theoretical Chemistry at University of Notre Dame. Beeper of boops.

GitHub Events

Total
  • Watch event: 5
  • Fork event: 1
Last Year
  • Watch event: 5
  • Fork event: 1

Committers

Last synced: 9 months ago

All Time
  • Total Commits: 1,238
  • Total Committers: 13
  • Avg Commits per committer: 95.231
  • Development Distribution Score (DDS): 0.605
Past Year
  • Commits: 0
  • Committers: 0
  • Avg Commits per committer: 0.0
  • Development Distribution Score (DDS): 0.0
Top Committers
Name Email Commits
John Parkhill j****l@g****m 489
jeherr j****r@g****m 402
Kun Yao k****o@n****u 259
David Toth d****1@n****u 51
Ryker McIntyre r****e@r****u 9
Your Name y****u@e****m 8
Ryker McIntyre m****r@g****m 7
Paolo Benigni p****i@g****m 4
ryker r****r@g****u 4
Thomas Heavey t****y@g****m 2
Axel Pahl a****l 1
jeherr j****1@n****u 1
John Edward Herr j****1@g****u 1
Committer Domains (Top 20 + Academic)
nd.edu: 3 garamon.crc.nd.edu: 1 gigan.chem.nd.edu: 1 rykers-mbp.dhcp.nd.edu: 1

Issues and Pull Requests

Last synced: 6 months ago

All Time
  • Total issues: 37
  • Total pull requests: 8
  • Average time to close issues: 19 days
  • Average time to close pull requests: about 3 hours
  • Total issue authors: 21
  • Total pull request authors: 5
  • Average comments per issue: 2.54
  • Average comments per pull request: 0.38
  • Merged pull requests: 6
  • Bot issues: 0
  • Bot pull requests: 0
Past Year
  • Issues: 0
  • Pull requests: 0
  • Average time to close issues: N/A
  • Average time to close pull requests: N/A
  • Issue authors: 0
  • Pull request authors: 0
  • Average comments per issue: 0
  • Average comments per pull request: 0
  • Merged pull requests: 0
  • Bot issues: 0
  • Bot pull requests: 0
View more stats
Top Authors
Issue Authors
  • jparkhill (11)
  • stefdoerr (3)
  • Dom1L (3)
  • xuwenshuo (2)
  • mkrompiec (2)
  • deleep225 (1)
  • PaoloBenigni (1)
  • xiexr151e (1)
  • nivedita104 (1)
  • Nevensky (1)
  • LiuTag (1)
  • naturalabundance (1)
  • Jussmith01 (1)
  • wangxr526 (1)
  • xuzhang5788 (1)
Pull Request Authors
  • AcylSilane (2)
  • theavey (2)
  • PaoloBenigni (2)
  • apahl (1)
  • jparkhill (1)
Top Labels
Issue Labels
enhancement (5)
Pull Request Labels

Packages

  • Total packages: 1
  • Total downloads:
    • pypi 5 last-month
  • Total dependent packages: 0
  • Total dependent repositories: 1
  • Total versions: 1
  • Total maintainers: 1
pypi.org: tensormol

TensorFlow+Molecules = TensorMol

  • Versions: 1
  • Dependent Packages: 0
  • Dependent Repositories: 1
  • Downloads: 5 Last month
Rankings
Stargazers count: 3.9%
Forks count: 5.1%
Dependent packages count: 7.3%
Dependent repos count: 22.1%
Average: 24.3%
Downloads: 82.9%
Maintainers (1)
jparkhill
Last synced: 6 months ago