Interpretable Machine Learning of Nanoparticle Stability through Topological Layer Embeddings Authors: Felipe Hawthorne, Leandro Seixas, James M. Almeida, Cristiano F. Woellner, Raphael M. Tromer Published: 2026 ABSTRACT The stability of chemically complex nanoparticles is governed by an immense configurational space arising from heterogeneous local atomic environments across surface and interior regions. Efficiently identifying low-energy configurations within this space remains a central challenge for first-principles–based materials discovery, particularly when the available reference data are limited. Here, we introduce a data-efficient and physically interpretable machine-learning framework based on a fragmented, layer-resolved descriptor that explicitly decomposes nanoparticles into surface, intermediate, and core environments using a topology-driven definition. This representation preserves a compact and fixed feature dimensionality while retaining spatial resolution, enabling controlled emphasis on different regions of the nanoparticle through physically motivated weighting schemes. Coupled with gradient-boosted decision-tree models and a ranking-based learning strategy, the proposed framework enables accurate identification of the most stable nanoparticle configurations using only a few hundred density functional theory reference calculations. Ranking performance metrics demonstrate near-saturation of correlation, high top-k recall, and rapidly vanishing regret at moderate training set sizes, highlighting the strong data efficiency of the approach. Beyond predictive performance, layer-weighting and SHAP-based interpretability analyses reveal how surface segregation, coordination topology, and local chemical disorder contribute differently to stability across spatial regions of the nanoparticle. The framework is system and code-agnostic, requiring only atomic coordinates, chemical species, and a scalar target energy, and is therefore directly transferable to other multicomponent nanostructures and to alternative first-principles or machine-learning energy methods. Introduction Understanding and predicting the stability of chemically complex nanoparticles remains a fundamental challenge in nanoscale materials science. − Recent advances in machine learning have further accelerated this field by enabling data-driven exploration and design of nanomaterials across vast structural and compositional spaces, complementing first-principles approaches in regimes where exhaustive sampling is computationally prohibitive. , Unlike bulk materials, nanoparticles exhibit pronounced heterogeneity in local atomic environments, arising from variations in coordination, chemical composition, and bonding across surface, subsurface, and core regions. These spatially distinct environments coexist within a single finite structure and give rise to a vast and rugged configurational energy landscape, even at fixed global stoichiometry. In multicomponent metallic nanoparticles, this complexity is further amplified by chemical disorder, , coordination heterogeneity, , and competing local bonding motifs that can differ markedly between surface and interior regions. , As a result, identifying low-energy configurations through direct first-principles sampling becomes rapidly intractable as the number of distinct atomic arrangements grows combinatorially with system size and chemical diversity. This intrinsic configurational complexity establishes a critical bottleneck for rational nanoparticle design and motivates the development of efficient, physically informed strategies for exploring the stability landscapes. A wide variety of structural descriptors have been proposed to encode atomic configurations for machine-learning models. − Atom-centered representations, such as symmetry functions, SOAP descriptors, , and graph-based embeddings, − have proven highly successful for learning local properties and constructing interatomic potentials. However, these approaches are typically high-dimensional and computationally demanding, and they often require large training data sets to achieve robust generalization. , As a consequence, their direct application to the efficient exploration of nanoparticle stability landscapes remains challenging in data-limited regimes. Recent studies have emphasized that the choice of structural descriptors is often the dominant factor governing data efficiency, transferability, and interpretability in machine-learning models for materials, frequently outweighing the impact of the specific learning algorithm employed. , Global descriptors provide an alternative strategy by summarizing entire nanoparticles or clusters through averaged geometric, topological, or chemical features. While these representations can improve data efficiency, they do so at the cost of spatial resolution, effectively mixing information from surface, subsurface, and core environments. In practice, this loss of spatial discrimination makes it difficult to disentangle region-specific contributions to stability and to rationalize how surface-driven effects compete with bulk-like energetic trends. , For nanoparticles, and particularly for chemically complex and core–shell–like systems, this lack of spatial resolution becomes especially limiting. Surface atoms typically dominate reactivity and often play a decisive role in stability, while interior atoms largely control cohesive energy, elastic response, and structural rigidity. In multicomponent alloys, variations in chemical short-range order, coordination topology, and local bonding motifs can differ markedly between surface and core regions, even when the global composition is fixed. As a consequence, stability cannot be reliably rationalized in terms of either purely surface-driven or purely bulk-like descriptors. Instead, it emerges from a coupled interplay between distinct spatial regions, whose relative energetic importance may vary across the configurational space. In this context, there is growing recognition that physically interpretable machine-learning models are essential not only for predictive accuracy but also for uncovering mechanistic insights and guiding rational materials design. Current descriptor frameworks rarely provide a systematic and interpretable way to isolate these regional contributions while maintaining a compact representation suitable for data-efficient learning, limiting their applicability for rational screening and design of complex nanoparticles. In this work, we introduce a fragmented, layer-resolved descriptor framework that explicitly decomposes nanoparticles into topological shells defined by their connectivity distances from the surface. By construction, this approach provides a physically motivated separation among surface, intermediate, and core environments without relying on arbitrary geometric cutoffs. The resulting representation preserves a compact and fixed feature dimensionality, independent of nanoparticle size or shape while retaining essential spatial resolution across distinct regions of the structure. This layered embedding enables controlled emphasis on different parts of the nanoparticle through physically motivated weighting schemes, allowing one to systematically probe how surface, subsurface, and core environments contribute to the stability. The descriptor is invariant to atom indexing, robust to small structural distortions, and directly applicable to nanoparticles with complex morphologies and chemical disorders. Importantly, it bridges the gap between fully local atom-centered descriptors and purely global representations by embedding spatial heterogeneity in a transparent, tunable, and interpretable manner. To directly address the practical objective of identifying the most stable nanoparticle configurations with minimal computational effort, we formulated the learning task as a ranking problem rather than a strict energy regression. This Perspective aligns naturally with high-throughput screening and active learning workflows, where the primary goal is to reliably down-select a small subset of low-energy candidates for expensive first-principles validation. Ranking-based formulations have been increasingly adopted in materials discovery workflows, particularly in active learning and high-throughput screening contexts, where relative ordering of candidates is more relevant than absolute property prediction. , Within this framework, we employ gradient-boosted decision-tree models, which are particularly well suited for data-limited regimes and structured, physically informed descriptors. While neural-network architectures can achieve high accuracy when trained on very large data sets, gradient boosting offers a favorable balance between expressive power, robustness against overfitting, and computational efficiency when only a few hundred reference calculations are available. , Finally, the combination of layer-weighted embeddings with post hoc interpretability tools enables a quantitative assessment of how different spatial regions of the nanoparticle govern the energetic stability. By linking model predictions to physically meaningful descriptors associated with surface, intermediate, and core environments, the proposed approach provides insights that go beyond black-box prediction and establishes a scalable, interpretable framework for accelerated discovery and analysis of stable nanostructures in complex multicomponent systems. We note that the present framework is complementary to machine-learning interatomic potentials (MLIPs). ,, MLIPs target the full Born–Oppenheimer surface and require energies and forces over typically 104–106 structures, after which they can perform structural relaxations directly. The present approach operates on already relaxed configurations, uses only scalar energies as labels, and reaches saturated ranking performance with a few hundred DFT-labeled structures, at the cost of being unable to relax structures on its own. An MLIP can generate the relaxed candidate pool, while the present ranking model selects the most stable members with minimal additional labels and full physical interpretability. Methodology First-principles calculations were carried out within the framework of Density Functional Theory (DFT) as implemented in the SIESTA code. The metallic nanoparticles were modeled from a icosahedric lattice placed inside a cubic supercell with dimensions of 35 × 35 × 35 Å, sufficiently large to prevent interactions between periodic images. Each species of the nanoparticle’s atoms was chosen at random, with replacements always respecting the chosen composition. Brillouin zone sampling was performed using a Monkhorst–Pack grid at Γ point. Electron–ion interactions were described using norm-conserving pseudopotentials, and the electronic wave functions were expanded in a localized atomic orbital basis set of double-ζ with polarization (DZP). Exchange–correlation effects were treated within the generalized gradient approximation (GGA) using the Perdew–Burke–Ernzerhof (PBE) functional. A real-space mesh cutoff of 250 Ry was employed. The Kohn–Sham equations were solved by direct diagonalization. Self-consistency was achieved using density matrix mixing with a weight of 0.035 and a convergence tolerance of 10–4. Spin polarization was not considered. Structural optimization was performed using the conjugate gradient method while keeping the supercell volume fixed. Atomic positions were relaxed until the maximum force on each atom was below 0.02 eV/Å, with a maximum atomic displacement of 0.1 Å per relaxation step. Atomic structures were analyzed through a graph-theoretical framework constructed directly from their atomic coordinates. Each structure consists of a set of N atoms with positions {r i } i = 1 N and chemical species {s i } i = 1 N . A chemical connectivity graph G = (V,E) was defined, where each vertex i ∈V corresponds to an atom and an undirected edge (i,j) ∈E exists if the interatomic distance d ij satisfies 1 dij≤λ(rinat+rjnat) where r nat is the ASE natural cutoff radius and λ is a global multiplicative factor. Distances were evaluated using the minimum image convention when periodic boundary conditions were present. This procedure provides an element-aware yet parameter-light definition of nearest-neighbor connectivity applicable to both crystalline and disordered systems. For each atom i, the coordination number was defined as the degree of the corresponding graph node 2 ki=∑jAij where A ij is the adjacency matrix of G. An ideal bulk coordination number k bulk was inferred automatically as the statistical mode of the {k i } distribution 3 kbulk=mode({ki}) under the assumption that bulk-like environments dominate the atomic population. Atoms with k i < k bulk were classified as surface seeds, reflecting the physical reduction of coordination at free surfaces and defects. To define surface and bulk regions in a robust and topology-driven manner, a breadth-first search was performed on G starting simultaneously from all surface seeds. This procedure assigns to each atom a topological distance li , defined as the minimum number of graph edges connecting atom i to any surface seed. Formally, 4 li=minj∈S⁡distG(i,j) where S denotes the set of surface seeds and dist G is the shortest-path distance on G. Atoms satisfying li