📄 chemistry

Physics-Guided Design of Ammonia Decomposition Catalysts with Interpretable Machine Learning

This study establishes a physics-guided screening framework that integrates density functional theory, microkinetic modeling, and interpretable machine learning to identify optimal NH* binding energy windows and discover high-performance Ru and Co catalysts for ammonia decomposition.

Original authors: Pragyansh Singh, Jiaqi Yang, Fanglin Che

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pragyansh Singh, Jiaqi Yang, Fanglin Che

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

Hydrogen is often hailed as the clean fuel of the future, capable of powering heavy industry and transport without carbon emissions. Yet, storing and moving hydrogen is notoriously difficult because it is a gas that takes up a lot of space and can be dangerous to handle. Ammonia offers a clever workaround. It is a liquid at relatively low pressures, easy to transport, and packed with hydrogen atoms. The challenge lies in getting that hydrogen back out efficiently. This requires a catalyst, a material that speeds up the chemical reaction to break ammonia apart into nitrogen and hydrogen without being consumed itself. While some metals can do this, finding the perfect one is like searching for a needle in a haystack. The surface of a catalyst is not a flat, uniform sheet; it is a complex landscape of hills, valleys, and atomic steps. The way atoms arrange themselves on these surfaces, combined with their electronic behavior, dictates how well they perform. Traditionally, scientists have had to test these possibilities one by one, a slow and expensive process that struggles to keep up with the vast number of potential combinations.

A team of researchers at Worcester Polytechnic Institute has developed a new way to navigate this complexity by combining deep physical understanding with modern artificial intelligence. Instead of trying to guess which catalyst might work, they first used detailed computer simulations to understand exactly what makes a catalyst fast. They focused on ruthenium, a metal known to be excellent at breaking down ammonia, and discovered that the speed of the reaction depends heavily on how tightly the intermediate pieces of the ammonia molecule stick to the surface. If they stick too tightly, the reaction gets stuck; if they stick too loosely, the reaction never starts. By modeling the entire process, they identified a specific "sweet spot" for this sticking strength where the reaction runs at its maximum speed. This target range became the goal for their search.

With a clear target in mind, the researchers built a massive library of data using quantum mechanics, the physics that governs how atoms interact. They calculated the behavior of over a thousand different configurations of ruthenium and hundreds of cobalt surfaces. These configurations varied in their shape, the number of extra atoms sitting on the surface, and their electrical charge. To make sense of this mountain of data without getting lost in the details, they trained a machine learning model. Unlike many artificial intelligence systems that act as black boxes, this model was designed to be interpretable, meaning the researchers could look inside and see exactly which factors were driving the predictions. They found that the model learned to rely on three main things: the electrical charge of the surface, the local arrangement of atoms where the reaction happens, and the overall energy of the surface structure.

The true test of this approach came when they tried to apply what they learned about ruthenium to cobalt, a different metal. Usually, a model trained on one metal fails miserably when applied to another because the atomic rules are slightly different. However, by feeding the model just a small amount of data specific to cobalt, they were able to adjust its understanding quickly. The model learned to transfer its knowledge, reducing its prediction errors significantly with very little new information. This suggests that the physical principles they identified are robust enough to work across different materials, provided the model is given a small nudge in the right direction.

Armed with these insights, the team generated a library of about thirty-three thousand new, physically possible catalyst structures. They used their trained model to screen these candidates, looking specifically for those that fell within the "sweet spot" they had identified earlier. The results were striking. By deliberately adjusting the electrical charge of the surfaces to be more negative, they were able to shift a large portion of their candidates into the desired performance range. In fact, more than half of the generated structures landed in the target zone, a success rate that was double what they would have achieved by chance. To ensure this wasn't just a computer trick, they selected nineteen of these promising candidates and re-calculated their properties using the rigorous, high-fidelity physics methods. The results held up: the model had correctly identified structures that would likely perform well, with only minor deviations.

This work demonstrates a powerful new path for designing materials. It shows that by combining the laws of physics with smart, transparent machine learning, scientists can bypass the need to test every single possibility. The researchers did not just find a better catalyst; they found a way to understand why certain structures work and how to steer the design process toward them. By focusing on the fundamental relationship between the surface's electronic state and its atomic geometry, they created a framework that can be adapted to other metals and reactions. This approach turns the search for new catalysts from a game of chance into a guided exploration, offering a clear roadmap for developing the efficient hydrogen technologies needed for a cleaner energy future.

Technical Summary: Physics-Guided Design of Ammonia Decomposition Catalysts with Interpretable Machine Learning

Problem Statement
Rational design of catalysts for ammonia (NH₃) decomposition is hindered by the vast configurational space arising from coupled structural and electronic effects. While Ruthenium (Ru) is a benchmark catalyst, its activity is highly sensitive to surface structure (e.g., stepped vs. flat facets) and the identity of kinetically relevant intermediates (N* vs. NH*). Exhaustive exploration of this space using first-principles Density Functional Theory (DFT) is computationally prohibitive, while experimental screening is impractical. Furthermore, traditional machine learning (ML) approaches often struggle with small datasets typical of catalysis or fail to preserve the geometric constraints inherent in physically realizable structures when using generative models. The core challenge is to establish quantitative, interpretable relationships between catalyst descriptors (structure, electronic state, coverage) and reaction energetics to efficiently screen for optimal binding energies without exhaustive DFT calculations.

Methodology
The authors propose a physics-guided, interpretable ML framework integrating three components: Microkinetic Modeling (MKM), DFT, and Machine Learning.

Microkinetic Modeling (MKM): The study first establishes a kinetic target. Using MKM on the most active Ru surface, Ru(10-13), the authors identify NH* dissociation as the rate-limiting step. They determine that a specific NH* binding energy ( $BENH$ ) window of -4.47 to -3.84 eV sustains near-maximal turnover frequency (TOF) across temperatures of 673–750 K and applied electric fields of 0 and -0.4 V/Å.
DFT Database Construction: A comprehensive database was generated containing 1,132 configurations for Ru and 270 for Cobalt (Co). These configurations span six surface facets, adatom coverages (0, 1/16, and 2/16 ML), various adsorption sites, and ten distinct charge states (ranging from +3 to -6 |e|) to simulate electronic perturbations analogous to applied electric fields.
Feature Engineering: Instead of raw atomic coordinates, the dataset was featurized using 12 physically motivated descriptors representing three regimes:
- Local binding environment: Coordination numbers (CN), generalized coordination numbers (GCN) of the site and adatoms, and adsorbate tilt angle ( $\theta_{NH}$ ).
- Adatom interactions: Inverse distance metrics (Prox) to handle cases with zero or multiple adatoms smoothly.
- Global surface state: Surface energy, adatom coverage, and explicit system charge.
Machine Learning Model: Five algorithms (Linear Regression, Ridge, SVM, Random Forest, XGBoost) were benchmarked. XGBoost was selected as the optimal model due to its ability to capture non-linear relationships, achieving a test Mean Absolute Error (MAE) of 0.082 eV on the Ru dataset.
Cross-Metal Adaptation: The framework was tested for transferability to Cobalt. By incorporating only 17% (45 configurations) of the Co data into the Ru training pool, the model adapted to Co chemistry, reducing the Co test MAE from 0.397 eV (zero-shot transfer) to 0.120 eV.
Interpretability and Screening: SHapley Additive exPlanations (SHAP) analysis was used to identify dominant descriptors. Guided by these insights, the authors generated a library of ~33,000 physically constructed candidate structures. These were screened using the trained model, with a specific focus on applying negative charge to weaken NH* binding.

Key Results

Descriptor Identification: SHAP analysis revealed that Charge, GCN of the binding site (GCNsite), and Surface Energy are the dominant factors governing NH* binding. Specifically, increasing negative charge (electron enrichment) systematically weakens NH* binding, shifting $BENH$ toward the target window.
Screening Efficiency: The physics-guided screening enriched the population within the target $BENH$ window. By biasing the screening toward negative charge states (-1 to -5 |e|), 50.7% of the generated candidates fell within the target window, compared to only 24.4% in an unbiased charge range.
Validation: A subset of 19 qualifying candidates underwent constrained-DFT validation (fixing NH* and adatom positions to match the screening geometry). The model predictions held with a mean absolute error of 0.172 eV. The validated DFT energies were slightly weaker than predicted (mean offset +0.104 eV), consistent with the lack of full geometric relaxation, but 12 of the 19 candidates remained within the target window.
Transferability: The framework demonstrated that a physically motivated descriptor set, trained primarily on Ru, can be efficiently adapted to Co with minimal additional data, suggesting the descriptors capture fundamental physical interactions rather than metal-specific artifacts.

Significance and Claims
The paper claims to establish a physics-guided design workflow that connects reaction kinetics, catalyst structure, and electronic state through an interpretable ML model. Its primary significance lies in:

Defining a Kinetically Informed Target: Moving beyond simple activity descriptors to define a robust $BENH$ window based on microkinetic analysis under realistic operating conditions.
Interpretable Forward Modeling: Proposing an alternative to generative inverse design for small-data regimes. By constructing physically realizable structures first and evaluating them with an interpretable forward model, the approach preserves geometric constraints and provides actionable design hypotheses (e.g., "increase electron density to weaken binding").
Cross-Metal Adaptability: Demonstrating that physically grounded descriptors allow for the transfer of learned structure-property relationships across different transition metals (Ru to Co) with limited data, reducing the computational cost of exploring new catalyst chemistries.
Scalability: Showing that this framework can navigate large configurational spaces (~33,000 candidates) to identify promising catalysts that would be difficult to find via random search or exhaustive DFT.

The authors conclude that this approach offers a generalizable strategy for translating reaction kinetics into experimentally actionable catalyst design principles, particularly for systems where data is scarce and geometric constraints are critical.

Technical Summary: Physics-Guided Design of Ammonia Decomposition Catalysts with Interpretable Machine Learning

More like this