Introduction

PiNN stands for pair-wise interaction neural network, a Python library built on top of TensorFlow for performing atomistic ML tasks. Besides training ML interatomic potentials, PiNN can also predict physical and chemical properties of molecules and materials (e.g. dipole moment and polarizability). It can be used together with the adaptive learn-on-the-fly plugin PiNNAcLe¹ and the heterogeneous electrode plugin PiNNwall² for modelling electrochemical systems.

Flexibility

PiNN is built with modularized components, and we try to make it as easy as possible. You do not have to rewrite everything if you just want to design a new network structure, or apply an existing network to new datasets or new properties.

Scalability

PiNN fully adheres to TensorFlow's high-level Estimator and Dataset API. It is straightforward to train and predict on different computing platforms (CPU, multiple-GPU, cloud, etc.) without explicit programming.

Examples

The quickest way to start with PiNN is to follow our example notebooks. The notebooks provide guides to train a simple ANN potential with a public dataset, import your own data or further customize the PiNN for your need.

Cite PiNN

If you find PiNN useful, welcome to cite it as:

[1] Li J, Knijff L, Zhang Z-Y, Andersson L, Zhang C. PiNN: equivariant neural network suite for modelling electrochemical systems. ChemRxiv. 2024; doi:10.26434/chemrxiv-2024-zfvrz This content is a preprint and has not been peer-reviewed.

[2] Y. Shao, M. Hellström, P. D. Mitev, L. Knijff, and C. Zhang. PiNN: a python library for building atomic neural networks of molecules and materials. J. Chem. Inf. Model., 60:1184–1193, January 2020. doi:10.1021/acs.jcim.9b00994.

Bibtex

@UNPUBLISHED{Li2024-wa,
  title    = "{PiNN}: equivariant neural network suite for modelling
              electrochemical systems",
  author   = "Li, Jichen and Knijff, Lisanne and Zhang, Zhan-Yun and Andersson,
              Linn{\'e}a and Zhang, Chao",
  abstract = "Electrochemical energy storage and conversion play an
              increasingly important role in electrification and sustainable
              development across the globe. A key challenge therein is to
              understand, control, and design electrochemical energy materials
              at atomistic precision. This requires inputs from molecular
              modelling powered by machine learning (ML) techniques. In this
              work, we have upgraded our pairwise interaction neural network
              Python package PiNN via introducing equivariant features to the
              PiNet2 architecture for fitting potential energy surfaces along
              with PiNet2-dipole for dipole and charge predictions as well as
              PiNet2-chi for generating atom-condensed charge response kernels.
              By benchmarking publicly accessible datasets of small molecules,
              crystalline materials, and liquid electrolytes, we found that the
              equivariant PiNet2 shows significant improvements over the
              original PiNet architecture and provides a state-of-the-art
              overall performance. Furthermore, leveraging on plug-ins such as
              PiNNAcLe for an adaptive learn-on-the-fly workflow in generating
              ML potentials and PiNNwall for modelling heterogeneous electrodes
              under external bias, we expect PiNN to serve as a versatile and
              high-performing ML-accelerated platform for molecular modelling
              of electrochemical systems.",
  journal  = "ChemRxiv",
  month    =  nov,
  year     =  2024,
  keywords = "Machine learning;Molecular dynamics;Liquid electrolyte;Ion
              transport;proton transfer;double layer;supercapacitor",
  language = "en"
}

@Article{2020_ShaoHellstroemEtAl,
  author    = {Yunqi Shao and Matti Hellström and Pavlin D. Mitev and Lisanne Knijff and Chao Zhang},
  journal   = {J. Chem. Inf. Model.},
  title     = {{PiNN}: A Python Library for Building Atomic Neural Networks of Molecules and Materials},
  year      = {2020},
  month     = {jan},
  number    = {3},
  pages     = {1184--1193},
  volume    = {60},
  doi       = {10.1021/acs.jcim.9b00994},
  publisher = {American Chemical Society ({ACS})},
}

¹ Y. Shao, and C. Zhang, “PiNNAcLe: Adaptive learn-on-the-fly algorithm for machine-learning potential,” arXiv:2409.08886, (2024). ↩
¹ T. Dufils, L. Knijff, Y. Shao, and C. Zhang, “PiNNwall: Heterogeneous electrode models from integrating machine learning and atomistic simulation,” J. Chem. Theory Comput. 19(15), 5199–5209 (2023). ↩