📄 chemistry

Identifying Hidden Trends in Complex Data to Design Best Metal-Organic-Framework Catalysts for Oxygen Evolution Reaction

This study introduces a "hidden trend mining" workflow that integrates machine learning with subgroup discovery to identify local structure-activity relationships in bimetallic thiophene-2,5-dicarboxylate metal-organic frameworks, successfully predicting and experimentally validating a highly efficient NiFe catalyst for the oxygen evolution reaction with an overpotential of 186 mV.

Original authors: Liangliang Xu, Jian Zhou, Canhui Zhang, Jiaqian Wang, Jinpei Huang, Zhong-Kang Han, Huawei Huang, Ning Xu, Heqing Jiang, Minghua Huang, Zhengxiao Guo, Sergey V. Levchenko, Xiaojuan Hu

Published 2026-09-01

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Liangliang Xu, Jian Zhou, Canhui Zhang, Jiaqian Wang, Jinpei Huang, Zhong-Kang Han, Huawei Huang, Ning Xu, Heqing Jiang, Minghua Huang, Zhengxiao Guo, Sergey V. Levchenko, Xiaojuan Hu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Making hydrogen fuel from water is a promising way to store clean energy, but the process is currently held back by a difficult chemical step called the oxygen evolution reaction. This is the moment when water molecules are split apart, releasing oxygen gas and leaving behind hydrogen. While the hydrogen is the valuable fuel, the reaction that produces the oxygen is slow and sluggish, requiring a significant amount of extra energy to get started. To make this process efficient, scientists need better catalysts—materials that speed up the reaction without being consumed themselves. For years, researchers have tried to design these catalysts by looking for simple rules that link a material's structure to its performance. However, in the complex world of chemical mixtures, these simple rules often fail because different combinations of atoms follow different physical laws, hiding the true patterns that lead to success.

A team of researchers has developed a new way to find these hidden patterns within a vast sea of chemical possibilities. They focused on a specific family of materials called metal-organic frameworks, which are like microscopic scaffolds built from metal atoms connected by organic chains. These materials are highly tunable, allowing scientists to swap different metals into the structure to see how it changes the catalyst's behavior. The researchers used a computer-based approach that combines artificial intelligence with chemical intuition to sift through hundreds of potential combinations. Instead of trying to force all the data into one single rule, their method identified and set aside the unusual cases that did not fit the main trend. By removing these outliers, the model could clearly see the local rules that govern the best-performing materials.

Using this strategy, the team screened more than 800 different metal combinations to find the most effective catalyst for the oxygen evolution reaction. The computer model pointed to a specific pairing of nickel and iron as the most promising candidate. To verify this prediction, the researchers synthesized the material in the lab and tested it in a basic solution. The results confirmed the computer's guess: the nickel-iron catalyst performed exceptionally well, requiring very little extra energy to drive the reaction and remaining stable for over 100 hours of continuous operation. This performance was significantly better than other tested combinations and even outperformed a standard benchmark material used in the field.

The study also revealed why this specific combination works so well. In these metal frameworks, the atoms are arranged in long, chain-like structures. The research showed that adding iron to the nickel framework changes how the material handles electrical charge. The iron acts as a reservoir that helps spread out the charge more evenly along the chain. This balanced distribution allows the material to hold onto the necessary intermediate chemical steps of the reaction without getting stuck or losing efficiency. This discovery explains why simple rules that work for some materials fail here: the ability of the surrounding structure to accommodate changes in electrical charge is just as important as the properties of the active metal itself.

By successfully separating the confusing data from the clear trends, this work demonstrates a new path for designing better catalysts. It shows that in complex chemical spaces, the best way to find a solution is not to look for a single universal rule, but to identify the specific conditions where a material behaves in a predictable and powerful way. The identification of the nickel-iron catalyst provides a concrete example of how this approach can lead to materials that make clean energy production more practical and efficient.

Technical Summary: Identifying Hidden Trends in Complex Data to Design Best Metal-Organic-Framework Catalysts for Oxygen Evolution Reaction

Problem Statement
Machine learning (ML) has become a potent tool for materials discovery; however, its efficacy is often limited in chemically heterogeneous spaces where a single global structure-property relationship fails to capture distinct physical rules governing different local environments. In catalysis, particularly for the Oxygen Evolution Reaction (OER), adsorption, redox, and multi-electron transfer processes are highly sensitive to local coordination and electronic structures. When diverse compositions are forced into a single global regression model, chemically meaningful local trends can be obscured by "out-of-trend" regimes that disproportionately increase model error. This challenge is acute in bimetallic Metal-Organic Frameworks (MOFs), where varying dopant metals alters local coordination, metal-oxygen covalency, and charge delocalization, potentially breaking simple scaling relations.

Methodology
The authors propose a "hidden trend mining" workflow designed to separate error-prone regimes from promising local trends within complex catalyst datasets. This approach integrates three core components:

Descriptor Learning: A set of 22 primary descriptors (including orbital energies, Bader charges, and electronegativities) was used. Symbolic regression (SISSO) was employed to derive a compact, physically meaningful descriptor, which was then incorporated as an additional feature.
Gradient Boosting Regression (GBR): A nonlinear regression model was trained to predict OER overpotentials.
Subgroup Discovery (SGD): Instead of classifying materials as active or inactive, SGD was used to identify and isolate "out-of-trend" compositions. The algorithm maximizes a quality function ( $Q$ ) defined by the reduction in 10-fold cross-validation mean absolute error (MAE) when a specific subgroup is removed from the dataset. This process iteratively refines both the subgroup selection and the GBR parameters to expose hidden local activity trends.

The workflow was applied to a dataset of 103 bimetallic thiophene-2,5-dicarboxylate MOFs ( $MaMb$ -TDC), constructed via Density Functional Theory (DFT) and active learning. The trained model was then used to screen over 800 candidate compositions.

Key Results

Model Performance: The integration of SGD with GBR and SISSO significantly improved predictive accuracy. The MAE for the left-out test set decreased from 0.25 V (GBR+SISSO) to 0.18 V. The identified "out-of-trend" subgroup consisted exclusively of compositions with high overpotentials (>0.5 V), effectively filtering noise from the low-overpotential region.
Catalyst Discovery: The model identified NiFe-TDC as a highly promising catalyst. Experimental validation across a series of $NiM_b$ -TDC catalysts (where $M_b$ = Fe, Co, Mn, Cu, Zn) confirmed the predicted trend.
Experimental Validation: NiFe-TDC demonstrated superior OER activity, requiring an overpotential of only 186 mV to reach 10 mA cm $^{-2}$ , outperforming NiCo-TDC (247 mV), Ni-TDC (249 mV), and the benchmark RuO $_2$ (264 mV). It also exhibited a low Tafel slope (51.5 mV dec $^{-1}$ ) and maintained stability for 100 hours at current densities of 10 and 100 mA cm $^{-2}$ .
Mechanistic Insight: Operando Raman spectroscopy and DFT calculations revealed that the active phase is a reconstructed NiFe oxyhydroxide. Electronic structure analysis identified "oxidation capacity" as the governing factor. Fe doping enhances charge delocalization within quasi-one-dimensional metal-oxygen chains, enabling the balanced stabilization of both OH* and O* intermediates. This mechanism explains why simple adsorption-energy scaling relations fail in this system; the dopant controls the framework's ability to accommodate oxidative charge redistribution.

Significance and Claims
The paper claims that hidden trend mining is a critical strategy for transforming heterogeneous catalyst data into interpretable descriptors. By separating error-prone regimes, the method exposes local activity trends that are masked in global models. The study demonstrates that this approach not only accelerates the screening of high-performance catalysts but also facilitates the discovery of chemically meaningful variables—specifically, the oxidation capacity of the metal-oxygen framework—that govern activity in complex bimetallic spaces. The authors position this work as a domain-specific, interpretable alternative to "black-box" global models, providing a framework for understanding how local microenvironments dictate catalytic performance in chemically diverse materials.

More like this