CRAFT (Chemical Recommendation & Analysis for Future Targets)

Jupyter notebook interface was made for an exploration and recommendation tool that identifies promising new compound compositions based on chemical similarity and site-specific elemental substitution. The notebook acts as a front-end to backend Python scripts that carry out advanced data processing and machine learning analysis.

How to run

pip install -r requirements.txt

run main.ipynb

How it works

1. Data Parsing and Grouping

   User selects what atomic sites are grouped based on the elements occupying them.

2. Machine Learning Analysis For each group of sites:

   • Partial Least Squares Discriminant Analysis (PLS-DA) is conducted using features derived from elemental properties.
   • Principal Component Analysis (PCA) is then performed on the loadings from PLS-DA to visualize chemical similarity in a lower-dimensional space.

3. Chemical Similarity Projection
   All remaining elements are projected into the PCA space, allowing the user to analyze and visualize chemical similarity between elements.

4. Visualization
   Each compound is visualized with atomic composition mapped to PCA-derived coordinates, representing the weighted contributions of the constituent elements.

5. Substitution Recommendation
   Based on the PCA space, the program recommends element substitutions under three modes:

   1. Fixed-site substitution — with some atomic sites fixed, find the best substitutes for the remaining sites.
   2. Alternative fixed-site combinations — explore other fixed-site combinations using known substitutions.
   3. Exploration within PCA-defined chemical space — identify promising elements within the learned chemical space.

6. Compound Generation
   After evaluating all recommendation modes, the most promising candidates are selected, and new compound compositions are proposed, considering stoichiometric constraints.