A Nonlinear Multi-Objective Prediction Strategy for Small-Sample Datasets in Homogeneous Catalysis

Area of Science:

• Catalysis
• Machine Learning
• Computational Chemistry

Background:

• Homogeneous catalytic reaction development faces challenges with resource-intensive experimental and computational methods, especially for small datasets and multi-objective optimization.
• Existing machine learning (ML) methods often require large datasets and struggle with small-sample data, high-dimensional spaces, and nonlinear relationships in catalysis.

Purpose of the Study:

• To present a nonlinear multi-objective ML workflow, PSO_CRP, for predicting reaction outcomes and performing interpretability analysis in homogeneous catalysis.
• To address the limitations of current ML approaches in handling small-sample datasets and complex catalytic processes.

Main Methods:

• Developed a Particle Swarm Optimization-based Catalysis Reaction Prediction (PSO_CRP) workflow.
• Utilized simple RDKit-derived molecular parameters, avoiding computationally expensive DFT calculations.
• Employed nonlinear multi-objective machine learning for predicting reaction categories and quantitative outcomes.

Main Results:

• The PSO_CRP workflow achieved higher predictive accuracy on four small-sample datasets compared to at least five common ML models.
• Key molecular descriptors influencing reaction outcomes were identified using permutation feature importance (PFI) and partial dependence plot (PDP) analyses.
• The identified descriptors and their contributions align with existing studies, offering enhanced mechanistic insight.

Conclusions:

• The PSO_CRP workflow provides a high-precision, low-cost, and interpretable framework for homogeneous catalysis.
• The approach offers valuable insights for forward prediction and rational catalyst design, particularly for small-sample scenarios.
• This method enhances model interpretability, guiding future catalyst development.