Training

ID4 Tutorials

These tutorials were developed by ID4 postdocs and graduate students.

Diversity-Aware Active Learning for Scientific Discovery

Author: Quan Nguyen

The ability to design informative, targeted experiments underpins many applications in scientific discovery. Active learning is a common tool for this task, using machine learning models to guide experimentation. This tutorial outlines recent efforts to endow active learning strategies with an emphasis on data diversity, which helps scientists avoid committing themselves to a single solutions that may be found to be infeasible, and overall tends to lead to better scientific understanding.

Slides

Geometric Algebra & Analysis: A New Paradigm for Scientific Computing

Author: Alex Guerra

This tutorial gives a short overview of geometric algebra (GA) and its associated calculus. Included are several applications, which give new computational tools for doing things like projection, differentiation, integration, and more on both euclidean and non-euclidean spaces. From this, we can do things like compute divergence & curl free fields, compute inside-outside oracles, and even compute topological features of the manifolds in consideration. We hope that this tutorial inspires people to consider using GA for modelling their own problems, as things which are difficult using standard vector calculus become simple with GA.

Slides

Graph Neural Networks in Deep Graph Library

Authors: Liyi Zhang, Adji Bousso Dieng

Deep Graph Library (DGL) is a scalable Python library for modeling graph neural networks and is framework agnostic, meaning that it integrates into frameworks such as PyTorch and TensorFlow. This tutorial will introduce basic data structures and semantics of DGL with line-by-line code. Then, it will show how to use DGL to build a graph convolutional network ready for training by any other auto-differentiation framework.

Slides    Python notebook

Molecular Dynamics with Allegro

Authors: Simon Batzner and Albert Musaelian, Anders Johansson, Lixin Sun, Boris Kozinsky

This is a tutorial for Allegro, an architecture for building highly accurate and scalable Machine Learning Interatomic Potentials (MLIPs). The goal of Allegro is to make it as simple as possible to train an acurate, fast, and scalable Machine Learning Interatomic Potential and deploy it in production simulations. You will never have to write a single line of Python, but instead you can train a network with a single command and easily use it to run MD in LAMMPS or ASE. If you need to customize it to your needs, the code is also modular and flexible under the hood.

Python notebook

JAX FDM

Authors: Rafael Pastrana, Sigrid Adriaenssens

This tutorial introduces JAX FDM, a differentiable form-finding solver powered by JAX for the automatic inverse design of 3D bar structures such as masonry domes, steel gridshells, and cable-net bridges. We provide an overview of the theoretical principles and challenges underlying the inverse design of 3D bar structures with the force density method (FDM), automatic differentiation, and gradient-based optimization; with mathematical notation and Python code.

The tutorial includes a demo of JAX FDM and an outlook of future opportunities to integrate JAX FDM within a data-driven, end-to-end pipeline.

Slides    Example 1 (code)     Example 2 (code)

Density Functional Theory (DFT)

Authors: Angela Pak, Elif Ertekin

This tutorial covers:

• an overview of the Schrodinger equation and the hierarchy of quantum chemistry/quantum physics techniques used when trying to solve it including wave function-based and Hamiltonian-based approaches
• a deeper dive into DFT, the workhorse method of computational materials science including: 1) its theoretical basis and most common formulations, 2) examples of what it is used for and where it struggles, and 3) basic workflows and some common codes that are available.

This tutorial also makes use of the Tiny DFT code, created by the Tiny DFT Development Team.

Slides   Tiny DFT on GitHub