Design Doc

[ncfrey]

Here we can plan out the structure of multipAL.

We should specify the following:

• Intent. What is the vision and intent for the package? Is it narrower or broader than "active learning for multi-property materials optimization"?
• Environment. What Python versions and imports are needed/supported?
• Classes in multipal/multipal.py. What classes do we need to write? What does each one do?
• Example notebooks. How will we show the functionality and what example datasets will be available?
• Tests. We should use test-driven development to build the necessary classes and use GitHub actions to automate testing.
• Interface. Will there be a CLI or is the package intended for use only programmatically in Python?
• Coding conventions. What code format, docstring conventions, linters, etc. will we use to keep the code uniform and readable?
• Docs. Will there be a website at readthedocs or somewhere else with documentation?

[ncfrey]

I'll put some initial thoughts/suggestions here and @jglazar feel free to change anything you like. The overall motivation is to keep the package as simple, usable, and low-maintenance as possible.

• Description could be: "MultipAL is a Python package for multi-objective materials optimization with active learning."
• Pick one version of python and stick to it, e.g. Python 3.7 or 3.8. Build a conda environment and all dependencies with pinned version numbers in requirements.txt.
• Classes can come from your jupyter notebook. Maybe an ActiveLearner class that takes a surrogate model and data can be the main class. We probably want to stick with single and joint-optimization to start with and add true Pareto frontier exploration later on.
• Add really small Jarvis and Materials Project datasets for tests and examples. It's good to keep overall repo size small so cloning and comparisons are fast.
• Tests can be tricky and annoying so after we plan out the classes, I'll write some bare bones tests as examples.
• Again keeping things simple, I think showing programmatic functionality with jupyter notebooks is good enough.
• I like yapf for formatting, numpy or Google docstring conventions, flake8 for linting. We can set up GitHub actions to automate these checks.
• Readthedocs is an easy way to automatically generate documentation on a free website.