Markov State Models for Tracking Reaction Dynamics on Catalytic Nanoparticles

By applying Markov state models to molecular dynamics simulations of hydrogen on rhodium catalysts, this study reveals that nanoparticle features and cooperative hydrogen interactions create complex, non-monotonic reaction kinetics that standard transition state theory fails to predict.

The Gist

Imagine you are trying to understand how a busy city park works. You want to know how people move, where they gather, and how they interact.

The Old Way (The "Static Map" Approach):
Traditionally, scientists studying chemical reactions (like how a catalyst helps turn hydrogen gas into something else) used a method called Transition State Theory (TST). Think of this like looking at a single, frozen snapshot of the park. You pick one specific spot, draw a line on the ground, and say, "If a person crosses this line, they are moving from 'sitting' to 'running'." You calculate how hard it is to cross that line and assume that's the whole story.

The problem? Real life isn't a frozen snapshot. The park is crowded, the ground is uneven, and people bump into each other. In chemistry, the "ground" is the surface of a catalyst (often a tiny metal nanoparticle), and it's constantly shifting, vibrating, and changing shape. The old method ignores all that chaos and assumes everything is perfectly still and empty except for the one person you are watching.

The New Way (The "Markov State Model" Approach):
This paper introduces a smarter tool called Markov State Models (MSMs). Instead of a frozen snapshot, imagine you have a high-speed drone camera filming the entire park for hours.

1. The Movie: You record the chaotic movement of thousands of people (atoms) bumping into each other, changing directions, and interacting with the environment.
2. The Pattern Recognition: You don't try to track every single step of every person. Instead, you use a computer to group the movie into "scenes."
   • Scene A: People hanging out on the smooth grass (flat surfaces).
   • Scene B: People stuck in the playground corners (edges and corners of the nanoparticle).
   • Scene C: People running fast in the middle of the crowd.
3. The Map: The MSM takes all that chaotic video and turns it into a simple flowchart. It tells you: "If you are in Scene B, you are likely to stay there for a long time before moving to Scene C." It identifies the slow parts of the process (the bottlenecks) without needing to know the exact path every single person took.

The Big Surprise: The "Trap" in the Corner

The researchers applied this "drone camera" method to Hydrogen interacting with Rhodium (a type of metal catalyst). They compared two shapes:

• Slabs: Flat, smooth metal sheets (like a flat park lawn).
• Nanoparticles: Tiny, bumpy balls of metal with sharp edges and corners (like a playground with slides and corners).

What they expected:
They thought the nanoparticles would be better catalysts. Why? Because corners and edges are usually "rougher" and more energetic, so they should grab hydrogen molecules and break them apart faster. It's like thinking a bumpy road makes a car go faster.

What they actually found:
The nanoparticles were actually slower at breaking hydrogen apart!

The Analogy:
Imagine the hydrogen molecules are like people trying to leave a crowded concert.

• On the Flat Slab: The crowd is spread out. People can easily find a path to the exit.
• On the Nanoparticle: The "corners and edges" act like sticky traps. When a hydrogen molecule gets stuck in a corner, it gets trapped there. It's like a person getting stuck in a narrow hallway or a corner of a room. They can't move easily, and they can't find the exit.

The study found that these "rough" corners and edges, which we usually think are helpful, actually slow down the reaction because they trap the hydrogen molecules, preventing them from reacting efficiently.

The "Crowded Room" Effect

The paper also discovered something weird about how crowded the surface gets:

• Low Crowding: If the surface is empty, adding more hydrogen helps the reaction (more people = more collisions).
• High Crowding: If the surface is too full, the reaction slows down.

The Analogy:
Think of a dance floor.

• If there are 5 people, they can't dance because they aren't bumping into each other.
• If there are 50 people, they bump into each other and dance (react) a lot.
• If there are 500 people, the floor is so packed that nobody can move. They are stuck bumping into each other without enough room to spin or jump. The reaction stops because the "dancers" (hydrogen atoms) are too crowded to find the momentum to react.

Why This Matters

This research is a game-changer because:

1. It sees the whole picture: It doesn't just look at one step; it watches the whole movie of the reaction, including the chaos and the crowd.
2. It fixes our intuition: It proved that "rougher" isn't always "better." Sometimes, the rough corners of a nanoparticle are actually holding the reaction back.
3. It helps design better catalysts: Now, engineers know that to make a better catalyst, they might need to design surfaces that avoid trapping molecules in corners, or they need to manage the "crowd" so the surface isn't too packed.

In short, the authors built a smart traffic control system for atoms. Instead of guessing how cars (atoms) move based on a static map, they watched the traffic flow in real-time and found that the "scenic routes" (corners and edges) were actually causing traffic jams.

Technical Summary

1. Problem Statement

Understanding the mechanisms driving catalytic selectivity and efficiency is critical for catalyst design. Traditional theoretical approaches rely on Transition State Theory (TST) combined with ab initio calculations to identify rate-limiting steps and energy barriers. However, these methods face significant limitations in heterogeneous catalysis:

• Static Assumptions: Standard TST typically assumes a static, defect-free surface and a single adsorbate, ignoring the dynamic restructuring of the catalyst and the fluctuating concentration of reactants under operating conditions.
• Combinatorial Explosion: As system complexity increases (e.g., nanoparticles with edges, corners, and defects), the number of possible atomic environments and elementary steps grows combinatorially. Manually identifying all processes and barriers becomes intractable.
• Dynamic Interplay: The interplay between surface restructuring, reactant diffusion, and reaction events occurs on comparable timescales, which static models cannot capture.

The authors propose that Markov State Models (MSMs) offer a solution to analyze complex dynamical data from Molecular Dynamics (MD) simulations without requiring a priori definitions of reactive events, thereby capturing the full dynamical complexity of catalytic systems.

2. Methodology

The study employs a workflow combining Machine Learning Interatomic Potentials (MLIPs), MD simulations, and MSM construction to analyze hydrogen ($H_2$) dissociation and association on Rhodium (Rh) catalysts.

• System Setup:

  • Catalysts: Four geometries were simulated: (100) and (111) Rh slabs, and 2 nm and 5 nm cuboctahedral nanoparticles.
  • Conditions: Simulations were run at $T = 450$ K with hydrogen coverages ranging from dilute (0.05 ML) to saturated (1.0 ML).
  • Potential: An Atomic Cluster Expansion (ACE) MLIP, fitted to DFT (PBE) data, was used to enable long-timescale MD simulations (accumulating 200 ns of trajectory data per system).

• Markov State Model Construction:

  1. Local Featurization: Instead of tracking global system changes, the authors used an atom-centered local featurization ($\phi_i$) for each hydrogen atom. This vector ($d_\phi = 131$) describes the local coordination environment (2-body, 3-body, and 4-body interactions with Rh and H neighbors within a 7 Å cutoff).
  2. Dimensionality Reduction: Time-lagged Independent Component Analysis (TICA) reduced the 131-dimensional feature space to 5 dynamically relevant dimensions ($d_f = 5$).
  3. Discretization: K-means++ clustering was performed in the TICA space to define 1,200 discrete microstates.
  4. Transition Matrix: A transition probability matrix $P(\tau)$ was estimated from the trajectory data for a specific lag time $\tau$.
  5. Kinetic Analysis: Eigenvalues and eigenvectors of $P$ were analyzed to identify slow dynamical modes, relaxation timescales, and equilibrium distributions.

• Rate Calculation:

  • Reactive fluxes were calculated using the committor function ($q^+$), estimated via the Dynamical Galerkin Approximation (DGA).
  • Reactant ($H_2$) and product ($H$) states were defined based on nearest-neighbor bond lengths, with intermediate states defined by specific clustering criteria to avoid artifacts from "trapped" top-site hydrogens.
  • Rates ($k_{AB}$) were derived from the net reactive flux, ensuring independence from lag time selection.

3. Key Contributions

• MSM Application to Catalysis: The paper successfully extends MSMs, traditionally used in biophysics (e.g., protein folding), to heterogeneous catalysis. It demonstrates a method to extract interpretable kinetic insights from high-dimensional MD data without pre-defining reaction pathways.
• Local vs. Global Dynamics: The authors introduce a "local environment" approach where the state of the system is defined by the local coordination of individual reactants rather than global system configurations. This allows for the tracking of specific reactive events (like $H_2$ dissociation) amidst complex surface dynamics.
• Identification of New States: The analysis revealed a previously unreported dynamically distinct state: $H_{top}$, where a hydrogen atom sits atop a Rh atom trapped by two neighboring hydrogens. While dynamically distinct, this state was found to have low population and minimal impact on the overall slow modes.
• Automated Mechanism Discovery: The approach automatically identifies the slowest reaction modes (e.g., diffusion between facets vs. association/dissociation) based on the spectral gap of the transition matrix, rather than relying on human intuition.

4. Key Results

• Counter-Intuitive Nanoparticle Behavior:
  • Contrary to the intuition that under-coordinated sites (edges/corners) on nanoparticles accelerate reactions, the study found that nanoparticles slow down $H_2$ dissociation compared to flat slabs.
  • Mechanism: Corners and edges act as dynamical traps for physisorbed $H_2$ molecules. The high energy barrier for dissociation from these sites, combined with the trapping effect, reduces the overall dissociation rate, particularly at high hydrogen concentrations.
• Non-Monotonic Rate Dependence:
  • Association Rates: The rate of $H \to \frac{1}{2}H_2$ association peaks at intermediate coverages and decreases at high saturation.
  • Cause: At high coverages, hydrogen atoms are crowded, leading to reduced mobility. While collisions are frequent, the atoms lack the necessary momentum to overcome the association barrier. This effect is more pronounced on (100) surfaces than (111) surfaces due to site availability.
  • Dissociation Rates: $H_2$ dissociation rates increase with concentration on slabs (due to momentum transfer and barrier lowering) but remain suppressed on nanoparticles due to trapping at edges/corners.
• Facet Dominance:
  • On nanoparticles, hydrogen tends to accumulate on (100) facets because binding there is energetically more favorable than on (111) facets. Consequently, the majority of the reaction flux occurs on (100) facets, even on nanoparticles.
  • The (111) surface allows for faster association at high concentrations compared to (100) because it offers more open hollow sites, reducing crowding.

5. Significance and Implications

• Beyond TST: The results highlight that standard TST fails to predict the non-monotonic behavior of reaction rates at high coverages and the specific trapping effects of nanoparticle geometries. TST assumes a static surface and independent events, missing the cooperative, dynamic interplay of adsorbates.
• Catalyst Design: The study suggests that maximizing catalytic activity for hydrogen dissociation may require under-saturating the surface or designing nanoparticles that minimize the trapping effect of edges and corners.
• Operando Conditions: The work emphasizes the necessity of simulating catalysts under realistic operating conditions (dynamic restructuring, varying coverage) to uncover unique nanoscale behaviors that static models miss.
• Future Outlook: The methodology provides a scalable framework for studying complex catalytic systems with multiple reactants. Future challenges include applying this to multi-element catalysts and integrating enhanced sampling techniques for systems with very high energy barriers.

In summary, this paper establishes Markov State Models as a powerful tool for bridging the gap between atomistic MD simulations and macroscopic catalytic kinetics, revealing that the "active sites" on nanoparticles can sometimes act as kinetic traps, fundamentally altering reaction dynamics in ways traditional theories cannot predict.