Competing adsorption of H and CO on Pd-alloy surfaces: Mechanistic insight into the mitigating effect of Cu on CO poisoning

This study employs a machine-learning-enhanced computational framework to reveal that while Au-rich Pd-Au-Cu surfaces suppress adsorption overall, the specific mitigation of CO poisoning by Cu arises from its ability to provide viable pathways for hydrogen absorption into the material when Pd-dominated paths are blocked.

The Gist

The Big Picture: The "Goldilocks" Hydrogen Sensor

Imagine you have a tiny, super-sensitive nose made of Palladium (Pd) metal. This nose is designed to sniff out Hydrogen gas ($H_2$) in the air. When it smells hydrogen, it changes shape slightly, which changes how it reflects light. This is how "plasmonic" hydrogen sensors work.

However, this metal nose has two big problems:

1. The "Sticky" Problem (Hysteresis): Once it smells hydrogen and swells up, it gets stuck. It doesn't shrink back down easily when the hydrogen is gone. It's like a sponge that swells with water but refuses to squeeze dry.
2. The "Bad Smell" Problem (CO Poisoning): Carbon Monoxide (CO) is a gas that smells terrible to our nose, but to this metal, it smells better than hydrogen. If even a tiny bit of CO is in the air, the metal ignores the hydrogen and grabs the CO instead. The sensor goes blind.

The Solution: Scientists tried mixing the Palladium with Gold (Au) and Copper (Cu) to fix these problems.

• Gold fixes the "Sticky" problem.
• Copper fixes the "Bad Smell" problem.

But here is the mystery: How does Copper actually stop the CO from taking over? Nobody knew the answer until this paper.

---

The Detective Work: A Digital Laboratory

To solve this, the researchers didn't just mix chemicals in a lab; they built a massive digital simulation.

Think of the surface of the metal as a crowded dance floor.

• The Dancers: Hydrogen atoms and CO molecules.
• The Floor: The Palladium, Gold, and Copper atoms.
• The Music: The temperature and gas pressure.

The problem is that the dance floor is chaotic. The dancers move, the floor rearranges itself, and sometimes the dancers get stuck in a bad formation. Calculating every single move with traditional computer methods is like trying to count every grain of sand on a beach while the tide is coming in—it takes too long.

The New Tool: The authors used a "Smart AI Coach" (Machine Learning).

1. First, they used a super-accurate but slow calculator (DFT) to teach the AI how the atoms behave.
2. Then, they let the AI run millions of simulations in seconds to see how the dance floor looks under different conditions (hot, cold, lots of hydrogen, lots of CO).

---

Discovery 1: The Preparation Matters (The "Warm-up" Routine)

The researchers found that how you prepare the metal matters more than what you mix into it.

Imagine a gym class.

• Scenario A (Cold Start): If you start the class in a room with no students (no Hydrogen), the Gold atoms (Au) rush to the front of the room (the surface) to get comfortable. This creates a "Gold-rich" surface.
  • Result: The Gold surface is very polite; it doesn't let Hydrogen or CO in. The sensor becomes useless because it can't smell anything.
• Scenario B (Hot Start): If you start the class with a room full of Hydrogen students, the Gold atoms get pushed to the back. The Palladium (Pd) stays at the front.
  • Result: The surface is now mostly Palladium. It loves Hydrogen. Even if CO tries to push in, the Hydrogen crowd is so strong that CO can't take over completely.

The Lesson: To make the sensor work, you must "warm it up" in a Hydrogen-rich environment first. This forces the surface to stay Palladium-rich, which is better at fighting off CO.

---

Discovery 2: The Copper Mystery (The "Secret Tunnel")

This is the most exciting part. The simulations showed that simply looking at who sits where on the surface (Thermodynamics) doesn't explain why Copper is so good at stopping CO poisoning.

In fact, the simulations showed that Copper and Gold look very similar on the surface regarding who sits where. So, why does Copper work so much better?

The Analogy: The Traffic Jam
Imagine the metal surface is a highway.

• Hydrogen is a car trying to get into a parking garage (the bulk of the metal).
• CO is a giant truck that parks in the most popular spots, blocking the entrance.

What happens with Gold?
Gold is like a roadblock. If a Hydrogen car tries to drive past a Gold atom, the road gets bumpy and hard to cross. The car gets stuck. Gold actually makes it harder for Hydrogen to get into the garage if the main road is blocked by CO.

What happens with Copper?
Copper is like a secret tunnel.
When CO trucks block the main highway (the surface), the Hydrogen cars are stuck. But, because Copper is there, it opens up a side door or a secret tunnel that the Hydrogen cars can use to slip into the garage, even while the CO trucks are blocking the front.

The "Aha!" Moment:
The paper concludes that Copper doesn't necessarily stop CO from parking on the surface. Instead, Copper provides a backup route. When CO blocks the "best" paths for Hydrogen to enter the metal, Copper creates "good enough" paths that are still open, allowing the sensor to keep working.

---

Summary: The Takeaway

1. Preparation is Key: You can't just mix the metals; you have to "train" the sensor by heating it in Hydrogen first. This forces the surface to be the right kind of metal to fight CO.
2. Gold vs. Copper:
   • Gold stops the sensor from getting "stuck" (hysteresis) but can make it harder for Hydrogen to enter if CO is present.
   • Copper is the hero against CO poisoning. It doesn't necessarily push CO away; instead, it acts as a bypass route, letting Hydrogen sneak into the metal even when CO is blocking the main door.

In everyday terms: If you want a hydrogen sensor that works in a dirty, smoggy world, don't just add a little bit of everything. Heat it up in clean air first, and make sure you have Copper in the mix to act as the "emergency exit" for the hydrogen when the bad guys (CO) try to block the door.

Technical Summary

1. Problem Statement

The research addresses critical limitations in Palladium (Pd)-based hydrogen sensors, specifically hysteresis (difference between absorption and desorption) and CO poisoning (preferential adsorption of CO over H, which blocks the sensor). While alloying Pd with Gold (Au) mitigates hysteresis and Copper (Cu) reduces CO poisoning, the underlying atomic mechanisms—particularly for Cu—are not fully understood.

• The Challenge: Modeling multi-component alloy surfaces (Pd-Au-Cu) under realistic operating conditions (finite temperature, reactive gas environments) is computationally prohibitive due to the vast configurational space involving surface segregation, adsorption, and thermodynamic effects.
• The Specific Mystery: Experimental evidence suggests Cu is crucial for mitigating CO poisoning, yet Cu rarely resides directly on the surface. The thermodynamic reasons for this effect are unclear, as standard adsorption models often fail to distinguish the specific benefit of Cu over Au.

2. Methodology

The authors developed a multi-scale computational framework to overcome the limitations of standard Density Functional Theory (DFT) while maintaining high accuracy.

• Machine-Learned Interatomic Potentials (MLIPs):
  • Trained on DFT data using two architectures: NEP (Neuroevolution Potential) for speed (active learning data generation) and MACE (Higher-order equivariant message passing neural networks) for high accuracy.
  • Used to generate a comprehensive training dataset of ~18,403 structures (586,000 atoms) covering bulk, surface, and adsorbate configurations.
• Cluster Expansions (CE):
  • Constructed for {111}, {110}, and {100} surface orientations.
  • The CE models describe the mixing energy of the metal lattice (Pd, Au, Cu) and adsorbate lattices (H, CO, vacancies).
  • Key Innovation: Used MLIPs to generate training data for the CE, effectively removing size limitations on the training set and implicitly including relaxation effects.
  • Employed pruning, merging, and regularization (LASSO + ARDR) to manage the massive parameter space (thousands of interaction parameters).
• Monte Carlo (MC) Simulations:
  • Performed in the semi-grand canonical ensemble (SGC) to simulate surface segregation and co-adsorption at specific temperatures (773 K for annealing, 300 K for operation).
  • Surface Phase Diagrams (SPD): Constructed continuous SPDs mapping H and CO coverage as a function of chemical potentials, rather than discrete points.
• Kinetic Analysis:
  • Used Nudged Elastic Band (NEB) calculations with the MACE model to determine migration barriers for H absorption into the bulk for dilute alloys.

3. Key Contributions

1. Computational Framework: Established an accurate and efficient workflow combining DFT, MLIPs (NEP/MACE), and Cluster Expansions to model complex multi-component surfaces with adsorbates under realistic conditions.
2. Continuous Surface Phase Diagrams: Demonstrated a method to generate continuous SPDs for H-CO co-adsorption, revealing how preparation conditions (annealing environment) dictate surface composition and subsequent adsorption behavior.
3. Mechanistic Insight into Cu: Provided a kinetic explanation for Cu's role in CO poisoning mitigation, distinguishing between thermodynamic adsorption preferences and kinetic absorption pathways.

4. Key Results

A. Influence of Preparation Conditions

• Surface Segregation: The surface composition is highly sensitive to the H coverage during annealing (preparation).
  • H-poor conditions: Au segregates to the surface, suppressing both H and CO adsorption (bad for sensing).
  • H-rich conditions: The surface becomes Pd-rich. These surfaces maintain higher H coverages compared to pure Pd at relevant CO partial pressures, indicating improved resistance to CO poisoning.
• Thermodynamics: Alloying generally increases the H-to-CO ratio compared to pure Pd, but this effect is largely independent of the specific ratio of Au to Cu in the bulk. Thermodynamics alone cannot explain why Cu is specifically required.

B. The Role of Cu vs. Au (Kinetic Barrier Analysis)

• Au Effect: While Au globally improves sorption kinetics in bulk alloys, locally, the presence of Au in the near-surface region creates high kinetic barriers for H absorption. Paths passing near Au atoms are energetically unfavorable.
• Cu Effect: Cu atoms in the surface region (even if not the topmost layer) leave the H absorption energetics unchanged compared to pure Pd.
• The Mechanism:
  • In pure Pd, CO blocks the surface almost entirely, preventing H absorption.
  • In Au-rich alloys, CO blocks the surface, and the remaining open sites near Au atoms have high kinetic barriers, hindering H entry.
  • In Cu-containing alloys, CO blocks the surface partially, but the remaining open sites (near Cu) offer viable kinetic pathways for H to shuttle into the bulk. Cu acts as a "gatekeeper" that remains permeable to H even when the most favorable Pd paths are blocked by CO.

C. Surface Orientation

• While {111} is the dominant surface, {110} and {100} orientations show similar trends.
• {110} surfaces show less benefit in the co-adsorption limit compared to {111}, and Cu segregation behavior on {110} differs (segregating more in vacuum), but the general conclusion regarding preparation conditions holds.

5. Significance

• Resolving the Cu Paradox: The study resolves the discrepancy between experimental observations (Cu is essential for CO resistance) and thermodynamic models (which show little difference between Au and Cu). It proves that the benefit of Cu is kinetic, not thermodynamic.
• Design Guidelines: The work provides a clear design strategy for hydrogen sensors:
  1. Prepare alloys in H-rich environments to ensure a Pd-rich surface.
  2. Include Cu in the alloy to provide kinetic pathways for H absorption when CO blocks the surface, rather than relying solely on Au which may kinetically hinder absorption near surface sites.
• Methodological Impact: The framework of using MLIPs to train Cluster Expansions offers a scalable solution for modeling complex, multi-component materials under reactive conditions, a task previously limited by computational cost.

In summary, the paper demonstrates that while thermodynamic tuning of surface composition via preparation conditions is crucial, the specific mitigation of CO poisoning by Copper is driven by its ability to maintain low kinetic barriers for hydrogen absorption in a CO-poisoned environment, a mechanism that Gold cannot provide.