As structure-property relationships in functional nanomaterials have been widely explored^[1,2], a common strategy is to identify the representative geometric structure that best explains observed material properties^[3-5]. This focus has generally motivated the identification of ground-state or low-energy structures^[2,6-8]. However, this paradigm does not fully capture the behavior of nanomaterials under realistic operating conditions. Nanomaterial surfaces and interfaces may continuously restructure in response to adsorbates^[9,10], electrolyte environments^[11], temperature^[12], support interactions^[13], and electrochemical potentials^[14]. Here, we use “reconstruction” to refer to such atomistic rearrangement under operating conditions. As a result, experimentally measured behavior often cannot be attributed to a single stable configuration but instead reflects contributions from condition-dependent ensembles of microstates.

The dynamic structural ensemble reframes the central question in nanomaterial reconstruction studies. The focus therefore shifts from a single stable structure to understanding how accessible states interconvert under operating conditions^[15,16]. Operando characterization provides essential information, but the measured signals are often indirect, time-averaged, making atomic-scale evolution difficult to interpret. For computational modeling, the challenge lies in identifying relevant structures within a vast and unknown configurational space that remains costly to explore at the first-principles level^[17]. Therefore, a practical approach should efficiently generate structures, sample metastable microstates, and validate structural distributions under realistic conditions^[18].

Recent advances in artificial intelligence (AI) and machine learning (ML) are opening new routes for exploring nanomaterial reconstruction. Generative models can learn structural patterns, propose candidate configurations, and sample complex distributions under physical, chemical, or experimental constraints^[19-22]. Related advances in AI-assisted structure generation, multiscale materials modeling, dynamic electrocatalyst simulations, generative materials design, and AI-Agent workflows^[23-28] further motivate integrated approaches to nanomaterial reconstruction.

In this Perspective, we consider reconstruction as an ensemble problem rather than the identification of a single representative structure. We discuss how generative models, combined with machine learning interatomic potentials (MLIPs), density functional theory (DFT), and operando measurements, can be used to explore dynamic structural ensembles under operating conditions. We further highlight the challenges in sampling, validation, and integration with experiment. Finally, we outline future opportunities for developing more reliable and physically grounded reconstruction workflows.

Nanomaterial surfaces under operating conditions cannot be fully represented by a single stable structure, as atomic rearrangements or even surface atom dissolution may cause multiple interconverting microstates^[29-31]. These microstates differ in defect structures, adsorption configurations, and local coordination environments. Their relative populations depend on different operating conditions^[3,10,30,32].

For example, Pt nanoparticles retain a clean surface under vacuum [Figure 1A], whereas CO adsorption induces high-index facets and the formation of new surface layers under CO conditions [Figure 1B]. Li et al. further theoretically revealed that even when Pt nanoparticles reach a dynamic structural equilibrium under CO conditions, their surfaces continue to undergo local structural fluctuations, including the transient formation and dispersion of small clusters [Figure 2]^[15]. These local structural changes are accompanied by fluctuations in catalytic activity, showing that the observed activity reflects an evolving structural ensemble rather than a single static morphology.

Theoretical reconstruction modeling can therefore be formulated as the inference of conditional probability distributions over metastable structures under operating conditions. The structural distribution can be described by P(X|c), where c denotes the relevant external conditions, such as temperature, pressure, electrode potential, pH, and others. Dynamic processes such as activation, deactivation, or phase transitions additionally require a distribution P(τ|c) over reconstruction pathways^[28,34].

We summarize the common generative model-based workflow integrating generative models, MLIPs, and DFT/ab initio molecular dynamics (AIMD) for reconstruction studies through three coupled functions: candidate generation, accelerated sampling, and validation. Generative models broaden the structural search space under operating conditions^[35,36]. MLIPs enable efficient sampling and high-throughput calculations^[18,37]. DFT/AIMD together with experimental constraints can refine and validate representative structures. This generate-accelerate-validate loop in Figure 3 delivers sets of (quasi-)steady-state structures rather than a single optimal configuration. Table 1 summarizes the main technical approaches discussed in this Perspective and their limitations.

After formulating reconstruction as the inference of a conditional structural distribution P(X|c), the practical challenge is to generate plausible structures under given operating conditions, infer their relative populations, and identify possible reconstruction pathways. Generative AI can contribute to these targets through structure generation, probabilistic ensemble inference, and sequence-based pathway exploration, as summarized in Figure 4A. Figure 4B schematically shows two representative generative schemes^[38-40,43]. Diffusion-based models progressively add noise to an initial structure x₀ through a fixed forward Markov chain and generate candidate structures through a learned reverse denoising process (indicated by dashed arrows)^[38-40,43].

Overall, future efforts on atomic-scale reconstruction of nanomaterial surfaces and interfaces should move beyond isolated structure screening toward condition-dependent structural ensembles and reconstruction pathways^[29,51,52], together with explicit uncertainty quantification. A central challenge is to represent operating conditions in a physically meaningful way. Temperature, pressure, chemical potential, solvent, electric-double-layer effects, and other external factors should be incorporated through physically appropriate representations or constraints. Such physics-grounded conditioning is critical for transferable ensemble prediction.

A practical route forward is to integrate conditional generative models^[46] with MLIP-accelerated dynamics^[18,37], first-principles and experimental validation^[29,48], and uncertainty-aware ensemble reweighting^[44,45]. Within this framework, agents based on large language models (LLMs) can integrate prior knowledge of nanomaterial design from the literature^[27], while multimodal models can integrate heterogeneous structural and chemical representations^[53]. Such agentic and multimodal systems can therefore serve as an orchestration and information-integration layer for physics-grounded structure generation and validation.

Current AI-assisted approaches to studying nanomaterial reconstruction still suffer from limited and biased training data, incomplete sampling of slow reconstruction processes, difficulty in simultaneously incorporating multiple physical and experimental constraints, and limited interpretability. Addressing these issues will require closer integration of thermodynamics, kinetics, and uncertainty quantification. Ultimately, the goal is not only to generate plausible nanostructures, but also to obtain evidence-consistent ensemble predictions that explain how nanomaterials restructure under realistic conditions and how such restructuring controls their functional properties.