Interpretable, Physics-Informed Learning Reveals Sulfur Adsorption and Poisoning Mechanisms in 13-Atom Icosahedra Nanoclusters

By combining dispersion-corrected density functional theory with physics-informed machine learning, this study elucidates the sulfur adsorption and poisoning mechanisms across 30 transition metal 13-atom icosahedral clusters, identifying the Ti-Zr-Hf isoelectronic triad as a balanced group for designing sulfur-tolerant subnanometer catalysts.

The Gist

Imagine tiny, spherical balls made of just 13 metal atoms. These aren't just any balls; they are like microscopic soccer balls (icosahedrons) that act as super-efficient workers in the world of chemistry, helping to speed up reactions. Scientists call these "nanoclusters."

However, these tiny workers have a major weakness: Sulfur. Think of sulfur as a sticky, toxic glue. When sulfur gets on these metal balls, it sticks so tightly that the balls stop working. This is called "poisoning," and it's a huge problem for making clean energy and chemicals.

The big question the researchers asked was: Which of these 13-atom metal balls can handle sulfur the best? Which ones get stuck, and which ones can keep working even when sulfur is around?

To answer this, the team used two powerful tools:

1. Super-Computer Simulations (DFT): Like a high-precision video game, they simulated how 30 different types of metal atoms behave when sulfur tries to stick to them.
2. Smart Pattern Recognition (Machine Learning): Instead of just looking at the numbers, they taught a computer to find hidden patterns and group the metals based on how they react to sulfur.

The Main Discoveries

1. The "Goldilocks" Zone
The researchers found that not all metals react the same way.

• Some metals are like Velcro: Sulfur sticks to them so hard that the metal ball gets distorted and breaks its shape. It's too strong.
• Some metals are like Teflon: Sulfur barely sticks at all. It's too weak to do any good work.
• The Winners: They found a special trio of metals—Titanium (Ti), Zirconium (Zr), and Hafnium (Hf). These three are like the "Goldilocks" of the group. Sulfur sticks to them firmly enough to do its job, but not so hard that it crushes the metal ball's structure. They are strong but flexible.

2. The "Stiffening" Effect
When sulfur lands on these metal balls, it's like a heavy backpack being put on a gymnast.

• For most metals, the gymnast (the metal ball) wobbles and changes shape significantly to carry the weight. This is bad because it changes how the ball works.
• For the winning trio (Ti, Zr, Hf), the gymnast absorbs the weight without losing their balance. The ball gets a little stiffer, but it stays in its perfect shape. The researchers measured this by "listening" to the vibrations of the atoms; the winning balls vibrated in a way that showed they were stable and strong.

3. The "Electronic Handshake"
The paper explains that the strength of the bond depends on an "electronic handshake" between the metal and the sulfur.

• The winning trio has just the right amount of electronic "give and take." They share electrons with sulfur effectively without getting overwhelmed.
• The researchers also tested what happens when a sulfur molecule (SO2) lands on these winners. The results confirmed that these specific metal balls are tough enough to handle sulfur without falling apart.

The Bottom Line

The scientists didn't just guess; they used a mix of detailed physics simulations and smart computer learning to map out exactly how 30 different metals react to sulfur.

They concluded that if you want to build a tiny, sulfur-resistant catalyst (a helper for chemical reactions) that won't get "poisoned" easily, you should look at the Titanium, Zirconium, and Hafnium family. These three form a special team that balances strength and stability better than any other metal tested in this study.

In short: They found the "superheroes" of the metal world that can fight off sulfur poisoning without losing their own shape.

Technical Summary

Technical Summary: Interpretable, Physics-Informed Learning of Sulfur Adsorption in 13-Atom Icosahedral Nanoclusters

Problem Statement
Transition-metal (TM) nanoclusters (NCs) in the subnanometer regime offer unique catalytic properties distinct from bulk materials due to quantum size effects, discrete electronic states, and high surface-to-volume ratios. However, their practical application in heterogeneous catalysis is frequently compromised by sulfur poisoning, where sulfur-containing species (e.g., H₂S, SO₂, SOₓ) strongly adsorb to active sites, leading to deactivation. While sulfur adsorption on extended surfaces is well-understood, the behavior on subnanometer clusters remains poorly explored due to pronounced finite-size effects, enhanced fluxionality, and low coordination numbers. There is a critical need to understand the atomistic mechanisms of S–NC interactions to design sulfur-tolerant nanocatalysts.

Methodology
The study employs an integrated framework combining dispersion-corrected Density Functional Theory (DFT) with interpretable Machine Learning (ML) to investigate sulfur adsorption on 13-atom icosahedral (ICO) nanoclusters (TM₁₃) spanning 30 transition metals (3d to 5d series).

• Computational Setup: Calculations were performed using spin-polarized DFT (VASP) with the PBE functional and empirical D3 dispersion corrections. Spin-orbit coupling (SOC) was tested for 4d and 5d elements. Systems were modeled in a 20 Å cubic box to avoid periodic interactions. Vibrational analyses (Hessian matrix) were conducted to confirm true energy minima and calculate Zero-Point Energy (ZPE).
• Descriptors: A comprehensive set of physically motivated descriptors was extracted, including:
  • Energetics: Adsorption energy ($E_{ads}$), interaction energy ($\Delta E_{int}$), and distortion energy ($\Delta E_{dis}$).
  • Structural: Average bond length ($d_{av}$), effective coordination number (ECN), and TM–S bond distances.
  • Electronic: d-band center ($\varepsilon_d$), HOMO-LUMO gap ($E_{gap}$), and Bader charge transfer.
  • Vibrational: Frequency spectra and ZPE-based descriptors reflecting lattice stiffening.
• Machine Learning Strategy:
  • Unsupervised Learning: K-means clustering and Principal Component Analysis (PCA) were used to map periodic trends and group metals based on their adsorption responses in descriptor space.
  • Supervised Learning: Regression models (ElasticNetCV, RidgeCV, and Explainable Boosting Machine) were trained to predict adsorption energies. A Leave-One-Feature-Out (LOFO) protocol was utilized to identify the most robust descriptors for generalization, ensuring model interpretability over raw predictive performance.
  • Validation: Selected clusters (Ti₁₃, Zr₁₃, Hf₁₃) identified by the ML framework underwent explicit first-principles calculations for molecular SO₂ adsorption using Ab Initio Molecular Dynamics (AIMD) and static DFT.

Key Contributions and Results

1. Systematic Mapping of Sulfur Adsorption: The study provides a comprehensive map of S adsorption on 30 different TM₁₃ ICO clusters. Results confirm that S adsorption is thermodynamically favorable across all metals, with adsorption energies ranging from 1.39 eV (Hg) to 11.70 eV (Mo).
2. Decomposition of Adsorption Mechanisms: The total adsorption energy is dominated by the interaction term ($\Delta E_{int}$), confirming chemisorption as the primary driver. However, distortion penalties ($\Delta E_{dis}$) vary significantly; while moderate for most metals, they are substantial for specific elements (e.g., Cr, Mo, W, Ir), indicating that S adsorption can induce significant structural rearrangements or lattice stiffening.
3. Periodic Trends and Electronic Structure:
   • Adsorption strength generally increases from 3d to 5d series due to increased d-orbital radial extent and relativistic effects enhancing d–p overlap.
   • The d-band center ($\varepsilon_d$) shifts to lower energies across the series, correlating with adsorption strength consistent with the d-band model.
   • Charge transfer from the NC to sulfur is observed in all cases, driven by electronegativity differences, with the largest transfer occurring in 3d clusters.
4. Identification of Resilient Isoelectronic Triad: Through unsupervised clustering and LOFO analysis, the study identifies the isoelectronic triad of Titanium (Ti), Zirconium (Zr), and Hafnium (Hf) as a distinct group. These clusters exhibit a "balanced" response: they bind sulfur strongly enough to facilitate activation but undergo minimal structural distortion, preserving the ICO motif. This behavior aligns with the Sabatier principle, suggesting an optimal regime for catalytic activity without severe poisoning-induced destabilization.
5. Validation with SO₂ Adsorption: Explicit calculations for SO₂ adsorption on the selected Ti₁₃, Zr₁₃, and Hf₁₃ clusters confirm strong binding and a tendency toward dissociation, linking the electronic states and lattice responses identified in the atomic S study to molecular poisoning mechanisms.

Significance and Claims
The paper claims that the integration of DFT with interpretable, physics-informed machine learning offers a robust pathway for screening nanocatalysts. By moving beyond "black box" predictions, the authors demonstrate that specific descriptors (interaction energy, distortion penalties, and electronic structure) can be used to rationalize periodic trends and identify chemically resilient materials.

The primary significance lies in the identification of the Ti–Zr–Hf triad as representative, sulfur-tolerant platforms. The study posits that these clusters offer a data-driven guideline for designing subnanometer catalysts that balance strong adsorbate activation with structural integrity, thereby mitigating the risks of sulfur poisoning. The work emphasizes that while ML accelerates screening, the physical interpretability of the models is essential for uncovering the mechanistic rules governing adsorption and poisoning in complex nanoscale systems.