Interpretable Machine Learning of Nanoparticle Stability through Topological Layer Embeddings

1. Problem Statement

The stability of chemically complex nanoparticles is governed by a vast configurational energy landscape arising from heterogeneous local atomic environments across surface, subsurface, and core regions.

• The Challenge: Identifying low-energy configurations via direct first-principles sampling (Density Functional Theory, DFT) is computationally prohibitive due to the combinatorial explosion of atomic arrangements.
• Limitations of Current ML: Existing machine learning approaches often rely on:
  • High-dimensional local descriptors (e.g., SOAP, symmetry functions) that require large datasets and lack spatial resolution.
  • Global descriptors that average properties over the entire particle, losing critical spatial information regarding surface vs. core contributions.
• The Gap: There is a need for a data-efficient, physically interpretable framework that preserves spatial resolution (distinguishing surface, intermediate, and core) while maintaining a compact feature dimensionality suitable for limited reference data.

2. Methodology

A. First-Principles Data Generation

• System: 1,000 distinct configurations of $Al_{70}Co_{10}Fe_{5}Ni_{10}Cu_{5}$ decagonal quasicrystalline alloy nanoparticles (55 atoms each).
• Method: DFT calculations performed using the SIESTA code with PBE-GGA functionals, norm-conserving pseudopotentials, and DZP basis sets.
• Goal: Generate a rugged energy landscape at fixed global stoichiometry to isolate the effects of atomic ordering and chemical disorder on stability.

B. Topological Layer-Resolved Descriptor

The core innovation is a fragmented, layer-resolved descriptor that decomposes nanoparticles based on graph theory rather than arbitrary geometric cutoffs.

1. Graph Construction: Atoms are nodes; edges exist if interatomic distance $d_{ij} \leq \lambda(r_i + r_j)$.
2. Layer Definition:
   • Surface Seeds: Atoms with coordination numbers below the bulk mode are identified as surface seeds.
   • Breadth-First Search (BFS): A BFS is performed from all surface seeds to assign a topological distance ($\ell_i$) to every atom.
   • Stratification: Atoms are partitioned into topological layers ($L=0$ to $L_{max}$) based on $\ell_i$. $L=0$ is the surface, $L=1$ is subsurface, and deeper layers represent the core.
3. Feature Extraction: For each layer $L$, the following features are computed:
   • Geometric: Mean coordination number, bond length statistics.
   • Topological: Ring fractions, Shannon entropy of cycles.
   • Chemical: Atomic fractions, Warren-Cowley short-range order parameters, chemical Shannon entropy, and element-pair bond statistics.
4. Weighted Representation: The final descriptor $D_w$ is a weighted sum of layer-specific features: $D_w = \sum w_L D(L)$. This allows researchers to probe the relative importance of specific regions (surface vs. core) by tuning weights $w_L$.

C. Machine Learning Strategy

• Task Formulation: The problem is framed as a ranking task (identifying the most stable configurations) rather than strict energy regression. This aligns better with high-throughput screening and active learning workflows.
• Model: XGBoost (Gradient-Boosted Decision Trees) was selected for its robustness in data-limited regimes and ability to handle structured, non-linear descriptors.
• Hyperparameter Optimization: Bayesian optimization (Optuna) was used to tune tree depth, learning rate, and regularization.
• Interpretability: SHAP (SHapley Additive exPlanations) values were used to quantify the contribution of specific features (e.g., surface Al fraction vs. core bond entropy) to the predicted ranking.

3. Key Results

A. Descriptor Analysis & Physical Insights

• Chemical Segregation: The topological layers revealed systematic segregation: the surface ($L=0$) is strongly enriched in Aluminum (Al), while transition metals (Co, Fe, Ni, Cu) are depleted. This trend reverses slightly in the core.
• Electronic Structure: Valence Electron Concentration (VEC) and electronegativity vary non-monotonically across layers, confirming that nanoparticles cannot be treated as uniform bulk systems even in their interiors.
• Energy Landscape: Despite fixed global composition, the 1,000 configurations span a wide energy range (~1.0 eV/atom), indicating that local ordering creates distinct stability basins.

B. Predictive Performance (Ranking)

• Data Efficiency: The model achieves near-saturation in performance with only 300–500 training samples (out of 1,000).
  • Spearman Correlation ($\rho$): Reaches >0.97.
  • Top-k Recall: High recall for identifying the true lowest-energy structures.
  • Regret: The energy difference between the model's top choice and the true global minimum vanishes rapidly as training size increases.
• Layer Weighting Experiments:
  • Uniform Weighting: Best overall generalization.
  • Surface-Emphasized: Good training fit but slightly reduced test generalization, indicating surface chemistry alone is insufficient.
  • Core-Emphasized: Surprisingly strong performance, suggesting core packing constraints are a dominant energetic driver, though not the sole factor.
  • Conclusion: Stability emerges from a coupled interplay between surface, intermediate, and core regions; no single region fully dictates stability.

C. Interpretability (SHAP Analysis)

• Uniform Model: Dominated by global chemical entropy and Al fraction, reflecting the importance of local mixing and segregation.
• Surface-Emphasized Model: Shifts importance to specific bond counts (e.g., Al-Al) and surface fractions, highlighting surface-specific motifs.
• Core-Emphasized Model: Highlights mean degree and global entropy, but surprisingly retains surface terms, proving that core stability is globally constrained by surface relaxation and stoichiometry.

4. Key Contributions

1. Topological Layer Embedding: A novel descriptor that explicitly separates nanoparticles into surface, intermediate, and core environments using a connectivity-based (topological) definition, avoiding arbitrary geometric cutoffs.
2. Data-Efficient Ranking Framework: Demonstrates that accurate identification of stable nanoparticle configurations requires only a few hundred DFT calculations when using gradient boosting and ranking-based metrics, rather than large-scale regression datasets.
3. Mechanistic Interpretability: By combining layer-weighting with SHAP analysis, the study provides a transparent physical narrative of how different spatial regions (surface segregation, core packing, chemical disorder) contribute to stability.
4. Active Learning Readiness: The framework is designed to integrate with active learning loops, allowing for the iterative generation and screening of new candidates with minimal computational cost.

5. Significance

This work bridges the gap between first-principles accuracy and data-driven exploration in nanomaterials science.

• Rational Design: It offers a pathway to rationally design complex multicomponent nanoparticles (e.g., for catalysis or sensing) by understanding how to manipulate surface vs. core chemistry to optimize stability.
• Scalability: The method is computationally efficient and scalable, making it applicable to other alloy systems and morphologies without retraining from scratch.
• Transparency: It moves beyond "black-box" predictions, providing materials scientists with actionable physical insights (e.g., "increasing surface Al fraction stabilizes the core") that can guide experimental synthesis.

The authors have made the codebase publicly available to facilitate reproducibility and adoption in the broader community.