Descriptor-based Machine Learning for Coordination-Engineered Single-Atom Catalysts

Status: Ongoing research.

Role: Student project leader.

Contribution: Responsible for data curation, feature and descriptor design, missing-data treatment, machine-learning modelling, candidate screening, workflow evaluation, and DFT-based mechanistic verification.

Research Overview

This project develops a machine-learning framework for investigating composition–coordination–performance relationships in single-atom catalysts for lithium–sulfur batteries. It converts dynamic cycling data into material-level performance metrics and integrates data preprocessing, missing-data treatment, symmetry-aware representation, feature selection, interpretable modelling, candidate screening, and DFT-based mechanistic verification.

The current objective is to construct physically meaningful descriptors that connect metal identity, local coordination environment, electronic properties, and electrochemical performance, while improving model robustness under heterogeneous and limited experimental data.

Current Work

1. Dynamic cycling-data processing:
   Automated extraction and transformation of cycling records into material-level performance indicators suitable for machine-learning analysis.

2. Coordination-aware representation and descriptor design:
   Construction and evaluation of representations that account for metal identity, elemental composition, local coordination environment, and relevant electronic or physicochemical properties.

3. Robust and interpretable machine learning:
   Missing-data treatment, feature selection, cross-validation, hyperparameter optimization, SHAP-based interpretation, and candidate ranking under limited-data conditions.

4. Mechanistic verification:
   DFT calculations are used to examine catalyst–polysulfide interactions and assess the physical plausibility of model-derived hypotheses.

Research Scope

Lithium–sulfur battery catalysts provide the primary application system for this project. The broader methodological goal is to understand when data-driven descriptors and validation workflows can be transferred to other electrochemical energy-storage materials without assuming universal transferability in advance.