TRI Hackathon 2020

This repository contains assets and code for the 2020 TRI hackathon focused on the development of sequential learning AI for materials science.

Installation

To set yourself up for the lessons, use the following installation procedure. We also recommend using Anaconda python with a fresh environment (e. g. conda create -n hackathon and conda activate hackathon or the like).

1. Clone the repo

   git clone https://github.com/TRI-AMDD/tri-hackathon-2020.git

   or if you're using ssh:

   [email protected]:TRI-AMDD/tri-hackathon-2020.git

2. Install requirements. Install numpy first so dependencies requiring numpy pre-install work.

   cd tri-hackathon-2020
   pip install numpy==1.18.3
   pip install -r requirements.txt

3. Install package (in development mode so changes in your local repo will be available).

   python setup.py develop

4. Start your jupyter server. Note that you may have to install a clean version of jupyter in your conda env in order to make sure your codes are using the correct dependencies and pathing.

   jupyter notebook
   

Tutorial

The tutorial will walk through an example of using CAMD for sequential learning. Note that using CAMD is not required to participate in the hackathon, but we encourage you to compartmentalize your logic into similar bounded objects, i. e. the AI into an "Agent" and any postprocessing into something like an "Analyzer". Also note that CAMD is an open-source code, so please feel free to issue PRs as needed.

Hackathon rules and objectives

The objective of the hackathon is to develop an effective sequential learning procedure for your chosen dataset that will progress in its sampling of your dataset in an efficient manner, i. e. will choose "good" candidates for experimentation.
Agents will be judged on the following merits:

1. Efficiency - does the SL process make "discoveries" quickly as it progresses?
2. Interpretability - can one understand what the SL is doing from a human perspective? were the lessons learned in the process of constructing the SL process valuable from a broader perspective?
3. Code - is the code well-constructed, modular, and able to applied elsewhere?
4. Creativity/X-factor - does the SL process contain elements that are cutting-edge, clever, or genuinely new ways to do ML/AI in materials science? This one will be highly subjective, of course.