Introduction to Machine Learning for Materials Science

Contact for feedback: machri@phys.au.dk

Lesson plan

1. Python for computational science intro / recap | Link
2. Pytorch / The Atomic Simulation Environment | Link
3. Potential energy surfaces | Link
4. Introduction to Machine Learning | Link
5. Machine Learning for atomic systems 1 | Link
6. Machine Learning for atomic systems 2 | Link

Installation

Colab installion

If you use Google colab, which is likely the easist way to get started, then installation is simple - at the top of the notebook just ensure that the line

!pip install git+https://github.com/Mads-PeterVC/imlms

Has been executed, then the required dependencies will have been installed in your session. If you using colab make sure to save a copy of each notebook to your Google drive or download it to your machine.

Local installation

If you want to your local machine, then I recommend creating a virtual environment and installing the dependencies in there.

Virtual environment

To create a virtual environment open a terminal - and then assuming you have Python installed (if not see here or here.) run the command

python3 -m venv <path/to/env>

Where you decide on <path/to/env>. Once the command is completed you can check that the environment has been created

ls <path/to/env/>

And then you can activate it

source <path/to/env>/bin/activate

Dependencies

We will be using a variety of packages so to make managing all of that easier I have made this repo a package too - installing it will install all of the required packages.

Installation can be done like so, in a terminal with an active virtual environment

pip install git+https://github.com/Mads-PeterVC/imlms

Goals

• Atomistic description of materials

  • Governed by Quantum Mechanics
    • Most commonly approximate using density functional theory.
  • Develop an understanding of potential energy surfaces.
    • Structural stability
    • Dynamics / Vibrational properties
    • Chemical reactions
  • Be able to use ASE to setup structures and start calculations.
    • Make a cluster / nanoparticle
    • Make a surface w. adsorbate.

• Implement a simple potential energy surface

  • Lennard Jones
  • Lennard Jones with forces from Pytorch
    • Requires some introduction to Pytorch
  • Local optimization
  • Global optimization

• Introduction to Machine Learning

  • Motivation
    • Reducing the computational cost of simulations by replacing expensive calculations with cheap ones, extends system size and time scales.
  • What is machine learning?
    • Categories
      • Supervised
      • Unsupervised
      • (Reinforcement learning)
    • Learning from data
      • What is 'data'?
        • How is it represented to a computer?
    • Algorithms
      • Nearest neighour classification
      • Curve fitting
      • Neural Networks
  • Practical
    • Sklearn: Excellent for off-the-shelf algorithms of many types.
    • Pytorch: for NN / Gradient based things

• Implement trainable machine learning potential

  • Understand descriptors / Invariance
    • Implement some Fingerprint or Behler-Parinello type descriptor.
  • Understand neural networks
    • Feed forward layers
    • Activation functions
    • Gradient Descent
    • Batching
  • Learning a potential
    • Global anzats
    • Local anzats