Spin Classification in Metal Complexes Using Data-Driven Techniques

High-performance computing (HPC) techniques for spin-state classification in inorganic complexes are computationally demanding. Recently, machine learning approaches have emerged as efficient alternatives for this task (ACS JPCA exemple).

This repository contains a Python program to classify spin states in first-row transition metal complexes using machine learning techniques.

---

1. Extracting and Analyzing Data

• Reads an input specifying:
  • Target metal ion
  • Spin states
  • Coordination number (4, 5, or 6)
• Builds a dataset 'distilling' OMOL25 dataset (~4 million structures)
• Converts molecular structures into SOAP descriptors using the DSCRIBE library
• Preprocesses data:
  • Normalizing SOAP data
  • Splitting into training and test sets
• Performs exploratory data analysis:
  • PCA
  • t-SNE
  • k-Means clustering
• Generates plots colored by spin state and dominant atom in the first coordination sphere
• Generates scree plot

2. SVM and Random Forest Benchmark

• Implements Support Vector Machine (SVM) classification:
  • Linear SVM
  • Kernel SVM
• Implements Random Forest classifier
• Generates confusion matrix plots and .txt file containing information about classification metrix

3. Multi-Layer Perceptron (MLP)

• Performs spin-state classification using an MLP
• Highly regularized architecture due to small dataset
• Can work on PCA-reduced data
• Supports class-balanced training (undersampling majority class)
• Generates confusion matrix plots and .txt file containing information about classification metrix
• Generates validation accuracy and loss plots

---

Usage

• install dependencies in requirements.txt
• Change input parameters in input.txt
• then simply: python3 main.py