CatalystML

Attempt-23 Physics-informed machine learning framework for predicting photocatalysts performance using chemically meaningful descriptors.

Highlights

• Built a synthetic Cu2O dataset calibrated to real published data (Tirumala et al., ACS Catalysis 2022).
• Trained and compared two models (Random Forest, XGBoost) on physically grounded descriptors.
• Used SHAP to confirm the models learned the real non-monotonic Mie-resonance relationship, not a spurious correlation.
• Built a recommendation step that predicted the optimal particle size within 3.4% of the paper's reported value, without being told the answer.

Scientific Basis

This dataset is scoped to Cu2O only. The particle-size dependence of photocatalytic activity is calibrated against real measured rate constants from:

Tirumala, R. T. A. et al. ACS Catalysis 2022, 12, 7975-7985. https://doi.org/10.1021/acscatal.2c00977

The paper reports a volcano-type relationship between particle size and photocatalytic rate, driven by dielectric Mie resonance rather than surface area. This dataset reproduces that relationship rather than assuming "smaller particles = higher activity."

Results

Model	R² Score	RMSE
Random Forest	0.962	0.02
XGBoost	0.951	0.02

The ML pipeline recovers a real, counterintuitive structure-activity relationship (bigger particles can outperform smaller ones due to optical resonance, not surface area) directly from data

Catalyst Recommendation by ML

Shape	Reference Peak	Random Forest	RF % Off	XGBoost	XGB % Off
Sphere	145 nm	140 nm	3.4%	140 nm	3.4%
Cube	215 nm	200 nm	7.1%	175 nm	18.8%

Model Explainability (SHAP)

Physics-Informed Machine Learning for Photocatalyst Discovery

Machine learning has become a powerful tool for accelerating materials discovery. However, many models behave like black boxes and provide little insight into why a material performs well.

This project explores a different approach.

Instead of relying only on statistical correlations, CatalystML uses chemically meaningful descriptors to predict photocatalytic activity and understand the scientific factors that control catalyst performance.

The motivation for this work comes from my research on semiconductor photocatalysts, particularly Cu₂O-based nanostructures, where optical properties, crystal structure, particle morphology, and surface chemistry strongly influence catalytic activity.

The long-term goal of this repository is to demonstrate how machine learning and materials science can work together to accelerate catalyst discovery while maintaining physical and chemical interpretability.

---

Project Goals

• Build machine learning models for catalyst performance prediction
• Use physically meaningful descriptors instead of arbitrary variables
• Interpret model predictions using Explainable AI (SHAP)
• Compare different machine learning algorithms
• Recommend promising catalyst candidates for future experiments

---

Planned Workflow

Experimental Data
        │
        ▼
Data Cleaning
        │
        ▼
Feature Engineering
        │
        ▼
Machine Learning Models
        │
        ▼
Explainable AI
        │
        ▼
Catalyst Recommendation

---

Repository Structure

CatalystML
│
├── data
├── notebooks
├── src
├── models
├── figures
├── results
├── docs
└── tests

---

Current Status

🚧 This repository is under active development. Attempt-23. Improved with new ideas, optimized and fixed bugs from previous attempts stored in a private drive.

The first version focuses on building an interpretable machine learning workflow for photocatalyst performance prediction.

---

About Me

I am a materials scientist working at the intersection of heterogeneous catalysis, photocatalysis, nanomaterials, and computational materials science.

My research interests include

• Photocatalysis
• Catalysis
• Machine Learning for Materials
• Materials Informatics
• Computational Materials Science
• Explainable AI

This repository documents my journey toward developing AI-assisted tools for catalyst discovery.