[Paper Review] Adaptive Catalyst Discovery Using Multicriteria Bayesian Optimization with Representation Learning

Jie Chen, Pengfei Ou+5 more|arXiv (Cornell University)|Apr 18, 2024

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes a multicriteria Bayesian optimization framework enhanced with uncertainty-aware representation learning to accelerate high-throughput catalyst discovery. By integrating density functional theory (DFT) with a novel atomistic neural network (UPNet) for automated feature extraction and uncertainty quantification, the method reduces required DFT calculations by 10× while achieving high prediction accuracy and enabling interpretable, multi-objective optimization for CO2 reduction reactions.

ABSTRACT

High-performance catalysts are crucial for sustainable energy conversion and human health. However, the discovery of catalysts faces challenges due to the absence of efficient approaches to navigating vast and high-dimensional structure and composition spaces. In this study, we propose a high-throughput computational catalyst screening approach integrating density functional theory (DFT) and Bayesian Optimization (BO). Within the BO framework, we propose an uncertainty-aware atomistic machine learning model, UPNet, which enables automated representation learning directly from high-dimensional catalyst structures and achieves principled uncertainty quantification. Utilizing a constrained expected improvement acquisition function, our BO framework simultaneously considers multiple evaluation criteria. Using the proposed methods, we explore catalyst discovery for the CO2 reduction reaction. The results demonstrate that our approach achieves high prediction accuracy, facilitates interpretable feature extraction, and enables multicriteria design optimization, leading to significant reduction of computing power and time (10x reduction of required DFT calculations) in high-performance catalyst discovery.

Motivation & Objective

• To address the challenge of navigating vast, high-dimensional catalyst composition and structure spaces in high-throughput screening.
• To develop a representation learning model that automatically extracts meaningful features from complex catalyst structures without manual descriptor engineering.
• To enable multicriteria optimization in catalyst discovery by simultaneously considering multiple performance metrics such as activity, selectivity, and stability.
• To reduce computational cost and time in DFT-based catalyst screening through an adaptive Bayesian optimization framework with uncertainty-aware acquisition functions.

Proposed method

• Introduces UPNet, a deep learning model that performs uncertainty-aware representation learning directly from atomic-scale catalyst structures using graph neural network architectures.
• Employs a constrained expected improvement (CEI) acquisition function within Bayesian optimization to balance exploration and exploitation across multiple evaluation criteria.
• Leverages density functional theory (DFT) calculations as the ground-truth evaluation for catalyst performance, integrated into a sequential optimization loop.
• Uses self-supervised pretraining on a large set of catalyst structures to improve generalization and reduce data requirements for downstream optimization.
• Applies uncertainty quantification in UPNet to guide active learning, prioritizing regions of high uncertainty for DFT evaluation.
• Integrates the entire pipeline into a closed-loop system that iteratively selects promising candidates for DFT validation, minimizing total computational cost.

Experimental results

Research questions

• RQ1 Can uncertainty-aware representation learning from atomic structures improve the accuracy and efficiency of catalyst property prediction in high-dimensional spaces?
• RQ2 How effectively can a multicriteria Bayesian optimization framework balance competing objectives such as activity, selectivity, and stability in catalyst design?
• RQ3 To what extent can representation learning reduce the number of required DFT calculations in catalyst discovery while maintaining prediction fidelity?
• RQ4 Can the proposed framework enable interpretable feature extraction that reveals meaningful structure-property relationships in catalyst materials?

Key findings

• The proposed method reduces the number of required DFT calculations by 10× compared to conventional screening approaches, significantly accelerating the discovery process.
• UPNet achieves high prediction accuracy in estimating catalyst properties, with uncertainty estimates that reliably guide the selection of informative candidates for DFT evaluation.
• The framework successfully identifies high-performing catalyst candidates for the CO2 reduction reaction across multiple criteria, including overpotential and Faradaic efficiency.
• Interpretable feature extraction reveals meaningful atomic-scale descriptors linked to catalytic performance, enhancing scientific insight into structure-property relationships.
• The constrained expected improvement acquisition function enables effective multi-objective optimization, balancing trade-offs between competing performance metrics.
• The method demonstrates robustness and generalization across diverse catalyst compositions, validating its applicability to complex materials discovery problems.