Active Learning for OER Alloy Catalyst Discovery

An active learning framework for efficiently navigating the chemical space of ternary alloy catalysts for the Oxygen Evolution Reaction (OER). This repository implements Gaussian Process Regression (GPR) with intelligent query strategies to minimize the number of expensive DFT calculations needed to identify promising catalyst compositions.

Table of Contents

• Overview
• Key Features
• Installation
• Quick Start
• Detailed Workflow
• File Descriptions
• Data Format
• Customization Guide
• License

---

Overview

The Problem

Discovering optimal alloy catalysts for OER requires screening vast compositional spaces. Traditional approaches require DFT calculations for every candidate composition, which is computationally prohibitive.

Our Solution

This active learning framework:

1. Generates a complete feature space representing all possible surface configurations
2. Intelligently selects which configurations to calculate via DFT using acquisition functions
3. Predicts energies for uncalculated configurations using Gaussian Process Regression
4. Iteratively improves predictions by selecting the most informative next calculations
5. Estimates catalyst activity using Boltzmann-weighted kinetic models

System Details

• Target system: Ni-Fe-Co ternary alloy catalysts
• Adsorbates: O and OH intermediates
• Surface sites: fcc and hcp hollow sites
• Feature space: 280,000 unique surface motifs
• Feature vector: 15-element fingerprint (3 metals × 5 coordination zones)

---

Key Features

✅ Comprehensive feature generation - Automated fingerprint creation for all possible surface configurations
✅ Intelligent sampling - Active learning acquisition functions to minimize DFT calculations
✅ Gaussian Process Regression - Uncertainty-aware energy predictions
✅ Periodic boundary handling - DFT-compatible dataspace generation
✅ Activity estimation - Boltzmann distribution-based catalytic activity calculation
✅ Extensible framework - Easily adaptable to other alloy systems and reactions

---

Installation

Prerequisites

• Python 3.8 or higher
• ASE (Atomic Simulation Environment) for structure handling

Required Dependencies

pip install numpy>=1.21.2 scipy>=1.7.1 pandas>=1.3.3 matplotlib>=3.4.3 scikit-learn>=0.24.2

Clone the Repository

git clone https://github.com/minhee2043/ActiveLearning_OER.git
cd ActiveLearning_OER

---

Quick Start

Complete Workflow (5 Steps)

# Step 1: Generate complete feature space (280,000 configurations)
python GPRdataspace.py

# Step 2: Generate DFT-compatible subset for calculations
python possibleFp.py

# Step 3: Run active learning cycles (GPR + acquisition function)
python mygaussian.py

# Step 4: Preprocess data for activity calculation
python sum_element.py

# Step 5: Calculate and visualize catalyst activity
python activity_plot.py

Converting DFT Structures to Features

For analyzing individual DFT-relaxed structures:

from motif_to_feature import Slab
import ase.io
import numpy as np

# Read DFT-relaxed structure
trajectory = ase.io.read('relaxed_structure.traj')

# Extract surface motif
motif = Slab(trajectory)

# Generate 15-element feature vector
# metals: reference order for counting
# zones: coordination shells around adsorption site
feature = np.array(motif.features(
    metals=['Ni', 'Fe', 'Co'],
    zones=['ens', 'sn', 'ssn', 'sf', 'ssf']
))

print(f"Feature vector: {feature}")
# Output: [Ni1,Fe1,Co1, Ni2,Fe2,Co2, Ni3,Fe3,Co3, Ni4,Fe4,Co4, Ni5,Fe5,Co5]

---

Detailed Workflow

Phase 1: Feature Space Generation

1.1 Complete Dataspace (GPRdataspace.py)

Generates all 280,000 possible surface configurations for the Ni-Fe-Co ternary system.

Input: None (uses default parameters: 3 metals, 5 zones)
Output: GPRdataspace.csv - 280,000 rows × 16 columns (15 features + 1 multiplicity)
Key parameters (modifiable in code):

nMetals = 3              # Number of metal types
zoneSizes = (3, 6, 3, 3, 3)  # Atoms in each coordination shell

1.2 DFT-Compatible Subset (possibleFp.py)

Filters the complete dataspace to only configurations compatible with periodic DFT calculations.

Input: GPRdataspace.csv
Output: possibleFp.csv - Subset of valid DFT configurations

---

Phase 2: Active Learning Cycles

2.1 Gaussian Process Regression (mygaussian.py)

Core active learning engine that iteratively selects calculations and builds predictive models.

Inputs:

• DFT training data (user-provided)
• possibleFp.csv (candidate pool)

Outputs:

• Trained GPR models for each iteration
• Predicted energies for all configurations
• Selected next calculation points from the candidate pool

---

Phase 3: Activity Calculation

3.1 Data Preprocessing (sum_element.py)

Aggregates features by total composition for activity estimation.

Input: Predicted energies for all configurations
Output: Composition-aggregated data

3.2 Activity Calculation (activity_plot.py)

Estimates catalytic activity using microkinetic modeling.

Input: Composition-aggregated energies
Output:

• Activity maps

Key equations:

Rate ∝ Σ_i w_i × exp(-ΔG_i / kT)
where w_i = multiplicity, ΔG_i = reaction free energy

---

File Descriptions

Core Scripts

File	Purpose	Input	Output
GPRdataspace.py	Generate complete feature space	Parameters	GPRdataspace.csv
possibleFp.py	Filter DFT-compatible configurations	GPRdataspace.csv	possibleFp.csv
motif_to_feature.py	Convert structures to features	ASE Atoms	15-element vector
mygaussian.py	Active learning engine	DFT data, possibleFp.csv	Predictions, next points
sum_element.py	Preprocess for activity	Predicted energies	Aggregated data
activity_plot.py	Calculate activity	Aggregated energies	Plots, rankings

Helper Modules

File	Purpose
helperMethods.py	Mathematical utilities (count_atoms, multiplicity, etc.)

---

Data Format

Input Data Formats

DFT Trajectory Files (.traj)

• Format: ASE trajectory format
• Required tags:
  • tag = 0: Adsorbate atoms (O, OH)
  • tag = 1: Surface layer atoms
  • tag = 2: Subsurface layer atoms
  • tag = 3: Third layer atoms
  • tag = 4: Bottom layer atoms
• Atoms: Must contain Ni, Fe, Co metals

Training Data CSV

Expected format for initial DFT data:

Ni1,Fe1,Co1,Ni2,Fe2,Co2,Ni3,Fe3,Co3,Ni4,Fe4,Co4,Ni5,Fe5,Co5,Energy
3,0,0,6,0,0,3,0,0,6,0,0,3,0,0,-4.52
2,1,0,5,1,0,2,1,0,5,1,0,2,1,0,-4.38
...

Output Data Formats

GPRdataspace.csv

# 280,000 rows × 16 columns
# Columns 1-15: Feature vector (metal counts in 5 zones)
# Column 16: Multiplicity (degeneracy factor)
3,0,0,6,0,0,3,0,0,6,0,0,3,0,0,1
2,1,0,6,0,0,3,0,0,6,0,0,3,0,0,3
...

possibleFp.csv

Subset of GPRdataspace.csv with only DFT-compatible configurations.

Prediction Output

Ni1,Fe1,Co1,Ni2,Fe2,Co2,Ni3,Fe3,Co3,Ni4,Fe4,Co4,Ni5,Fe5,Co5,Predicted_Energy,Uncertainty
3,0,0,6,0,0,3,0,0,6,0,0,3,0,0,-4.50,0.15
...

---

Customization Guide

For Different Alloy Systems

Change Metal Types

File: GPRdataspace.py, motif_to_feature.py

# Original: Ni-Fe-Co system
metals = ['Ni', 'Fe', 'Co']

# Example: Pt-Pd-Rh system
metals = ['Pt', 'Pd', 'Rh']

Change Number of Metals

File: GPRdataspace.py

# Original: ternary system
nMetals = 3
range(3)  # in combinations_with_replacement

# Example: quaternary system (4 metals)
nMetals = 4
range(4)  # update this line

Modify Coordination Zones

File: GPRdataspace.py

# Original: 5 zones with (3,6,3,3,3) atoms
zoneSizes = (3, 6, 3, 3, 3)

# Example: 3 zones for simpler fingerprint
zoneSizes = (1, 6, 3)  # 1 on-top + 6 neighbors + 3 subsurface

File: motif_to_feature.py

# Update zone list accordingly
zones = ['ens', 's', 'ss']  # for 3-zone system

For Different Adsorption Sites

On-Top Adsorption

File: motif_to_feature.py

# Automatic detection
is_ontop = motif.onTop()
feature = motif.features(['Ni','Fe','Co'], onTop=is_ontop, zones=['ens','s','ss'])

# Manual specification
feature = motif.features(['Ni','Fe','Co'], onTop=True, zones=['ens','s','ss'])

Different Zone Definitions

Create custom zone methods in Slab class:

def custom_zone(self, onTop=False):
    '''Your custom coordination zone definition'''
    if onTop:
        return self.closest(layer=1, start=X, stop=Y)[0]
    else:
        return self.closest(layer=1, start=A, stop=B)[0]

For Different Acquisition Functions

File: mygaussian.py

# Upper Confidence Bound (default)
def acquisition_ucb(mu, sigma, kappa=2.0):
    return mu + kappa * sigma

# Expected Improvement
def acquisition_ei(mu, sigma, y_best):
    Z = (mu - y_best) / sigma
    return (mu - y_best) * norm.cdf(Z) + sigma * norm.pdf(Z)

# Probability of Improvement
def acquisition_pi(mu, sigma, y_best):
    Z = (mu - y_best) / sigma
    return norm.cdf(Z)

For Different Kernels

File: mygaussian.py

from sklearn.gaussian_process.kernels import RBF, Matern, RationalQuadratic

# Original: RBF kernel
kernel = RBF(length_scale=1.0) + WhiteKernel(noise_level=0.1)

# Alternative: Matern kernel (more flexible)
kernel = Matern(length_scale=1.0, nu=2.5) + WhiteKernel(noise_level=0.1)

# Alternative: Rational Quadratic (scale-invariant)
kernel = RationalQuadratic(length_scale=1.0, alpha=0.1) + WhiteKernel(noise_level=0.1)

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Contact

For questions, issues, or collaboration:

• GitHub Issues: https://github.com/minhee2043/ActiveLearning_OER/issues
• Email: [email protected]