A presentation given and written by Christopher Ostrouchov all contributions are welcome. We will be using the materials project and it's available data to "predict" material properties through machine learning. Many of the examples may be trivial but the focus is on introducing the workflow that is typical in machine learning.
The goal of this set of notebooks is to introduce you to the most important concepts of machine learning. While there are many many algorithms for fitting your data the methodology of gathering, sanitizing, investigating, and evaluating the goodness of fit is mostly the same. I hope to show you the process along with showing some methods from each branch of machine learning. Python has evolved into a great solution for easily performning these steps and along with R are great choices. My favorite description of Python is that it is the 2nd best language for every problem. Also it is probably the best glue language out there.
Python is a language that while it comes with "batteries included" most of the functionality is provided through packages. I myself may consider myself an "expert" the standard library (packages that are included by default with python) but there are always new packages to learn. The packages that we will be using:
-
requests for gathering the materials project data
-
pandas for storing data, sanitizing, and investigating the data. A supercharged excell spreadsheet.
-
matplotlib visualizing data
-
numpy used underneath the covers for pandas and basis of linear algebra in python
-
scikit-learn most popular machine learning library. Does not perform neural network calculations. Overview of available algorithms (does not cover all) link
-
pymatgen a package by the Materials Project for working with material science structures and analysis of calculations.
These packages have many many features but learning these core libraries will be more than enough for getting started.
Resources that we will be using that are not python specific are:
-
mybinder which is a way to make a custom programming environment available for free hosted on google cloud. Note that resources are limited about 1 CPU and 8 GiB RAM per instance. It is awesome you should use it too.
-
materialsproject based at Lawrence Berkely National Lab using the NERSC resources they provide data from their VASP similations available using a RESTfull API.
To get started we will lanch the introduction notebook with binderhub.
All contributions, bug reports, bug fixes, documentation improvements, enhancements and ideas are welcome! These should be submitted at the [Gitlab repository](https://gitlab.com/costrouc/ mse-machinelearning-notebooks). Github is only used for visibility.
Contributors:
- Chris Ostrouchov (maintainer)
MIT