Sustainability-Aware Discovery of Dual-Atom Hydrogen Evolution Catalysts via a Unified Machine Learning Framework for Activity and Carbon Cost
summary
The gist
The paper presents an accelerated discovery method using machine learning potentials to screen nitrogen-coordinated dual-atom catalysts for the hydrogen evolution reaction…
In short
A unified machine learning framework screens 460 dual-atom catalysts for hydrogen evolution reaction activity and monetized carbon cost simultaneously. The model predicts catalytic performance at DFT accuracy and converts embodied CO2 into a monetary burden. This method efficiently selects highly active catalysts with low environmental impact, such as Mn2@2Nf, by balancing performance against carbon expense.
Key concepts
- Machine Learning Potential (MLP)
- An advanced AI model trained on vast amounts of Density Functional Theory (DFT) data to predict the energy and forces of a catalyst's atomic structure. This allows researchers to simulate complex chemical reactions much faster than traditional physics-based calculations, achieving near-DFT accuracy for screening thousands of potential materials.
- Hydrogen Evolution Reaction (HER)
- The chemical process where water is split into hydrogen gas and oxygen gas, which is a key step in producing green hydrogen. The paper focuses on catalysts that can speed up this reaction efficiently under specific conditions to make sustainable fuel production possible.
- Carbon Cost Monetization
- A method used to assign a financial value to the carbon dioxide emitted when manufacturing a catalyst. By converting the embodied CO2 of each candidate into a monetary cost, researchers can select materials that offer high performance while minimizing their overall environmental footprint.
Terminology used across episodes
This episode discusses
- Sustainability-Aware Discovery of Dual-Atom Hydrogen Evolution Catalysts via a Unified Machine Learning Framework for Activity and Carbon Cost · Paper Radio
- Bayesian E(3)-Equivariant Interatomic Potential with Iterative Restratification of Many-body Message Passing
The paper
Sustainability-Aware Discovery of Dual-Atom Hydrogen Evolution Catalysts via a Unified Machine Learning Framework for Activity and Carbon Cost · Read on arXiv
Yanmei Zang, Hyun Gyu Park, Gi Beom Sim, Tae Hyeon Park, Ho Jin Lee, Xiaorong Zou, D. ChangMo Yang, Soohaeng Yoo Willow, Áe Jung Kim
Department of Energy Science, Sungkyunkwan University Department of Energy, Center for 2D Quantum Heterostructures Institute for Basic Science
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Sustainability-Aware Discovery of Dual-Atom Hydrogen Evolution Catalysts via a Unified Machine Learning Framework for Activity and Carbon Cost".
Mira: Here's the gist of the paper. The paper presents an accelerated discovery method using machine learning potentials to screen nitrogen-coordinated dual-atom catalysts for the hydrogen evolution reaction, which is crucial for sustainable hydrogen production The core finding of this work is that intermediate coordination environments (2N–4N) generally exhibit the most favorable catalytic performance in TM2@Nx–Gr systems The study systematically screened 460 distinct dual-atom configurations across various transition metals and nitrogen coordination environments to map the HER landscape Mapping the Catalytic Landscape The researchers investigated 23 transition metals across 20 nitrogen coordination motifs to map the HER landscape of graphene supported TM2@Nx-Gr DACs They found that intermediate coordination (2N to 4N) consistently yields near-optimal ΔGH∗, with specific candidates emerging as standout materials These standout candidates include Ti2@2Na, Mn2@2Na, Fe2@3Na, and several others exhibiting metallic or narrow gap (<0.25 eV) character The screening established how nitrogen coordination number and local geometry regulate charge transfer and hydrogen adsorption thermodynamics in TM2@Nx-Gr systems Machine Learning Potential Development A machine learning potential (MLP) was developed using the RACE, an E(3) equivariant graph neural network potential, to model the adsorption thermodynamics of TM2@Nx-Gr catalysts across all 460 configurations This MLP was trained on a DFT dataset comprising 26,406 unique configurations to reproduce near DFT level energies and forces at a fraction of the cost The training process utilized the Adam optimizer and minimized a weighted Huber loss of energy and force residuals Validation and Accuracy The RACE MLP demonstrated robust structural optimization capability for both TM2@Nx-Gr and H–TM2@Nx-Gr systems across different nitrogen coordination environments The model achieved excellent agreement with the corresponding DFT-relaxed geometries, confirming its ability to capture local atomic configurations accurately Correlation between ML predictions and DFT calculations for binding energy showed a high R2 value of 0.997 and very low errors (MAE = 17 meV) Gibbs free energy predictions also exhibited a strong correlation with DFT results, yielding an R2 value of 0.961 Catalytic Performance Trends The volcano plots of the exchange current density (i0) as a function of ΔGH∗ showed a well-defined volcano-type dependence on the hydrogen adsorption free energy (ΔGH∗), in accordance with Nørskov’s model The highest catalytic activity is achieved near thermoneutral hydrogen binding (ΔGH∗ ≈ 0 eV) Increasing nitrogen coordination induces a systematic modulation of ΔGH∗, reflecting the progressive tuning of the electronic structure of the diatomic metal centers Systems with intermediate nitrogen coordination (2N–4N) exhibit a higher density of active sites located near the volcano apex Electronic Structure Insights The analysis revealed that most TM2@Nx-Gr systems possess metallic or narrow-band-gap semiconducting characteristics, which are highly favorable for rapid electron transport during electrocatalysis Specifically, 40 configurations were identified as metals, while 11 configurations were classified as semiconductors with relatively narrow band gaps generally below 0.25 eV This metallic or narrow-gap semiconducting behavior supports the potential for high HER activity by facilitating rapid electron-transfer kinetics The developed RACE machine-learning framework, implemented within the BAMtorch framework, enables rapid and reliable exploration of large configurational spaces with near-DFT level accuracy This approach significantly reduces the computational cost associated with conventional first-principles calculations for catalyst screening The work establishes nitrogen coordination engineering as an effective strategy for optimizing dual-atom HER electrocatalysts Furthermore, it highlights the potential of the RACE machine-learning potential as a powerful and computationally efficient platform for next-generation catalyst discovery Computational Methods Spin-polarized DFT calculations were performed using VASP to optimize structures and evaluate electronic properties The RPBE(U)+TS approach was adopted for the 2D systems considered, utilizing the Tkatchenko–Scheffler (TS) dispersion correction The Gibbs free energy for the adsorption of H atoms (ΔGH∗) was calculated using the computational standard hydrogen electrode (SHE) model The exchange current density at pH = 0 was calculated using Nørskov’s model Machine Learning Method A machine learning interatomic potential was trained on a DFT dataset comprising 26,406 unique configurations using the equivariant graph neural network RACE as implemented in BAM-torch Atomic environments were encoded as graphs with a 6.0 Å neighbour cutoff and three interaction layers Training was carried out on NVIDIA A100 80 GB GPU hardware using the Adam optimizer and minimized a weighted Huber loss of energy and force residuals The learning rate was initialized at 1 × 10−3 and reduced by a factor of 0.
Paper summary: Kai: So we’ve seen that they’re looking at graphene supported transition metal two atom catalysts, or TM2@Nx-Gr DACs. The core finding they highlight is that intermediate coordination environments, specifically two to four nitrogen atoms, generally give the best catalytic performance in terms of hydrogen adsorption energy.
Mira: That makes sense because those intermediate environments seem to tune the electronic structure just right for the reaction. They found specific candidates like Ti2@2Na and Mn2@2Na standing out as strong performers <ref:2605.29821#pg1>.
Lev: So they’re saying that if you’re designing these systems, focusing on those two to four coordination numbers is a good starting point for finding high activity.
Kai: But the real power of this paper isn't just the activity; it’s how they factored in the carbon cost. They take all four hundred sixty configurations and compare them based on both their performance and their embodied carbon.
Mira: They showed that even though some materials have similar catalytic activity, like those based on Manganese or Iron, there is a huge range in their carbon cost.
Lev: That’s what the abstract points out: the most active candidates span metals like Manganese, Iron, Zirconium, Chromium up to Rhodium and Gold.
Kai: They found that some of these best performing catalysts have carbon costs that are tens of thousands of times lower than others, depending on how you look at it.
Mira: It’s about optimizing the trade-off between performance and environmental impact simultaneously within this unified machine learning framework.
Conclusion: Kai: Looking at the full paper, this work is about using machine learning to map out the landscape of hydrogen evolution catalysts while simultaneously accounting for their environmental footprint.
Mira: The authors are Yanmei Zang, Hyun Gyu Park, Gi Beom Sim, Tae Hyeon Park, Ho Jin Lee, Xiaorong Zou, D. ChangMo Yang, Soohaeng Yoo Willow and Hye Jung Kim.
Lev: And the implication is that this unified approach helps us discover materials that aren't just good catalysts but are also economically viable for real-world sustainable hydrogen production.
Kai: They use the machine learning potential to make screening much faster than traditional density functional theory calculations could manage across all those configurations.
Mira: So, in simple terms, this paper suggests that we can use these ML tools to find materials where you get high activity and low carbon cost at the same time without having to test every single possibility with expensive first-principles calculations.
Lev: It moves the search from being a brute force DFT exercise to a smarter, more targeted discovery process.
Kai: That’s what they are proposing: using this unified machine learning framework to accelerate the discovery of these dual atom catalysts.
More episodes
- 2610.32678-Probing the limits of materials design via a comprehensive crystal database of 14.6 million structures with 140 properties
- 2610.33595-Native-oxide-driven layer asymmetry, two-step switching and exchange bias in Co/Ir/Co trilayers
- 2503.19572-Spin models from nonlinear cellular automata
- 2606.23350-Irreversibility Enhances Quantum-Enhanced Markov-Chain Monte Carlo
- 2608.02744-Two-Parameter Ansatz for the Violation of Eigenstate Thermalization
- 2210.00806-Low- and high-energy localization landscapes for tight-binding Hamiltonians in two-dimensional lattices
- 2409.09625-Selective switch for band-edge alignments in pentagonal CdSeTe monolayer: Atom-valley locking
- 2606.20168-Norms, overlaps and Yangian descendants for the Haldane-Shastry spin chain
- 2512.19503-Strain-induced splitting of the CCDW-NCCDW phase transition in 1T-TaS 2
- 2307.12638-Strong localization of microwaves beyond two dimensions in aperiodic Vogel spirals