unsupervised_periodic_table

This is a repo that contains all the information to represent a periodic table of elements with machine learning tools, mainly unsupervised learning.

Project Organization

├── README.md          <- The top-level README file for understanding this 
│                         project.
├── data
│   ├── raw            <- The original data dump.
│   ├── interim        <- Data transformed but not as a final step.
│   └── processed      <- Data used to train the final model. 
│
├── models             <- Folder with the models and relevant information.
│
├── notebooks          <- Contains jupyter notebooks or related files.
│
├── references         <- Information relevant to the project context.
│
├── results            <- Folder with the obtained results.
│
├── src                <- Python code to run this project (main package).
│   ├── data           <- Scripts for data read and transformations. 
│   └── model          <- Scripts for ML workflow.
│
├── general_pipeline.py  <- Steps to run and execute the main task of this
│                           repo from data download to process and creation
│                           of a model with corresponding results.
│
├── setup.py           <- The requirements file for reproducing the
│                         analysis environment.
│
└── requirements.txt   <- Makes project pip installable.

Data

The data of this analysis is obtained from the PubChem webpage, which is an official website of the United States government. In this cite, one can find structured and machine-readable information on the Chemical Group Block, Atomic Mass, Standard State, Electronic Configuation, among other relevant fields (just to name a fiew). The next figure shows an example on the information that the periodic table of this cite could contain.

As relevant part of context, the periodic table as we know it today is managed by the IUPAC, and this is relevant since PubChem is working with IUPAC to make this information available and downloadable.

In the src/data/download_data.py script, one can see how to obtain the csv file from this webpage.

In this case, the main purpose of this analysis is to determine if machine learning techniques (mainly within an unsupervised scope) can reproduce some of the patterns that are seen in the periodic table by considering the new available structured information that has been integrated to each element over the years.

To do this a complete data science workflow was implemented, from de download and cleaning of the data, to the feature engineering, additional process and further transformations of the information to finally develop some unsupervised models and as well some interpretability (with shap values).

How to execute the code

In this case, the execution to replicate the solution by implementing the following code is fairly simple (I recommend to make use of an empty virtual environment when attempting to install the dependencies).

The first step, after cloning the repo, is to pip install the dependencies found in the requirements.txt file with:

pip install -r requirements.txt.

Then it is as simple as executing the following command to download, process the information and generate an unsupervised model with the corresponding shap explicability:

python general_pipeline.py

This should be everything that must be done to execute the code. This will as well yield the results as images in the results folder and the relevant models and information will be saved in the models folder.

A fast glance to some results

What one might find is that there are some elements that are somehow linked together. These elements are sub-groups and they are somehow intertwined with each other.

An example of this is the following image.

In this case, we can see that the atomic numbers are: [3, 4, 11, 12, 20]. Which correspond to Litium (Li), Beryllium(Be), Sodium (Na), Magnesium (Mg) and Calcium (Ca) respectively. In this specific instance Be, Mg and Ca belong to the same Element Group, and Li and Na belong to their own corresponding Group too. These elements can be found together in the Periodic Table of Elements which can also tell us that they have relatively similar fields.

The main reasons of similitud between these Elements can be found in the following figure which considers the shap values.

In this case, the most relevant features are (in order of importance):

1. Electron configuration highest orbital (is it s,p,d or f).
2. Electron configuration information count (after noble gas).
3. Electron configuration last orbital
4. Atomic radius over the atomic number (considering overlap).
5. Electron configuration electron hold.

Additional Considerations

This project is executed with Python 3.8.10.

Generated by Daniel Hernández Mota.

Link to github

References

• National Center for Biotechnology Information (2023). Periodic Table of Elements. Retrieved November 6, 2023 from https://pubchem.ncbi.nlm.nih.gov/periodic-table/.
• Vettigli, G. (2018) {MiniSom: minimalistic and NumPy-based implementation of the Self Organizing Map. Retrieved November 6, 2023 from https://github.com/JustGlowing/minisom/
• Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst. 30, 4768–4777 (2017).