Smarter Data: Rethinking Data Generation for Machine Learning Potentials in Heterogeneous Catalysis Authors: Wenbo Xie, Yixiao Han, Chenyu Wu, P. Hu Published: 2026 ABSTRACT Machine-learning potentials (MLPs) promise near-first-principles fidelity at scales relevant to heterogeneous catalysis, yet the key determinant of their reliability remains the quality of the training data. A paramount challenge lies in constructing training sets that capture both the system-specific structural domains and the broader chemical variability arising from the complex, dynamic nature of heterogeneous catalytic systems. In this perspective, we argue that the key question is no longer simply how to generate more data, but how to generate the right data. We reinterpret the main strategies used to construct MLP training sets for heterogeneous catalysis through the lens of scope, relevance, and coverage. We then examine the challenges of data generation in catalytic systems and how they have shaped prevailing structure-based sampling practices, which encode system relevance and targeted coverage directly into the data set. We also discuss emerging interaction-based sampling strategies that aim to broaden local-interaction support beyond narrowly predefined systems. We conclude by consolidating a possible hybrid data generation workflow that combines the strengths of both approaches, thereby bringing MLPs closer to the simulation of complex heterogeneous catalytic systems. Introduction Data lies at the heart of any machine learning model, whether in large language models, computer vision models, or, as considered here, machine learning potentials (MLPs) for chemistry. High-fidelity MLPs are rapidly reshaping atomistic simulation by offering near-first-principles accuracy at much larger scales. In heterogeneous catalysis, this promise is particularly attractive. − Catalytic reactivity is governed by dynamic reaction systems in which catalyst structure and composition, surface coverage, and reaction pathways all evolve with reaction conditions. Yet this same complexity also makes heterogeneous catalysis one of the most demanding application domains for MLPs, as catalytic interfaces are rarely confined to a narrow and well-defined configuration space. As a result, the central bottleneck in catalytic MLP development is often not model architecture alone, but whether the training data adequately represent the environments that the model will actually encounter during deployment. , This point is easy to overlook because much of the recent progress in atomistic machine learning has been framed in terms of architecture: better descriptors, message-passing schemes, and increasingly powerful equivariant neural networks. , These advances have been essential, but they do not remove a simpler and more fundamental constraint: a model can only interpolate reliably within the domain defined by the configurations and labels it has seen. In heterogeneous catalysis, this data challenge is amplified by the difference between idealized models and operando reality, a gap that we are seeking to bridge. Capturing this breadth of relevant environments and chemistry efficiently remains a challenge even for MLPs. At present, MLPs often focus on one specific aspect of a given catalytic system, for example, structural changes of the catalyst, , surface coverage, , metal-support interaction, and liquid/solid interface. , Correspondingly, many successful MLPs in catalysis are built as highly targeted models, designed for a particular catalyst, adsorbate, or mechanistic question. The sampling process is often wrapped in active-learning loops while uncertainty and diversity criteria select configurations for DFT labeling. This strategy is often highly effective because it narrows the domain and allows the training set to be tailored closely to the intended task. However, the success of this system-specific paradigm also reveals its limitations. As the field moves toward more realistic catalytic problems, the cost of constructing a new data set and retraining a new MLP for each catalyst, condition, or mechanistic regime becomes increasingly difficult to sustain. The issue is not that targeted models are misguided; on the contrary, they will remain indispensable for many well-defined applications. Rather, the challenge is that heterogeneous catalysis is increasingly demanding models that are not only accurate within a narrow domain, but also more robust to evolving environments, coupled complexities, and previously unseen local chemistries. This creates growing interest in broader and more transferable MLPs, whether through larger public data sets, − foundation model pretraining, − or more interaction-centered sampling strategies. , At the same time, these developments raise a deeper question: what actually makes a training data set good for heterogeneous catalysis? Put differently, if current workflows differ mainly in how they spend expensive first-principles labels, how should data set quality be judged, and what kind of data are most worth generating? This perspective argues that the key question is no longer simply how to generate more data, but how to define the right data for MLPs, as the language of data set design has remained comparatively underdeveloped. Here, we propose a common framework for thinking about data set design in heterogeneous catalysis, organized around three dimensions: scope, relevance, and coverage. These dimensions, we argue, provide a practical way to navigate the competing objectives inherent to data generation in heterogeneous catalysis. We first define what they mean in the context of MLPs and then revisit current data-construction strategies through this lens, including structure-based workflows, adaptive expansion and active learning, and more recent efforts to broaden support beyond narrowly predefined systems. Our aim is not to advocate a single universal solution, but to clarify the trade-offs underlying current practice and to show how future progress may come from combining complementary data generation philosophies to achieve both task-specific reliability and broader chemical generality. More broadly, best practices for data generation in heterogeneous catalytic MLPs are yet to come. What We Mean by “Smarter Data” In our view, a useful way to think about data set design begins with scope. Before asking how broadly to sample, one must first ask what scientific question the model is meant to answer. Is the goal to describe catalyst restructuring, coverage effects, support interactions, long-time dynamics, or explicit reaction events? Different questions imply different target distributions, and scope defines both the purpose of the MLP and the initial logic of data set construction. Once the scope is defined, the next question is relevance. The issue is not simply how much data to add, but what kind of data must be included for the model to remain reliable for the intended task. This is especially important in heterogeneous catalysis, where the key challenge is often to identify what is actually important for the problem at hand. For example, if the goal is to model metal-support interaction, the data set should not only contain supported metal configurations, but also sufficiently sample the metal and support subsystems individually, so that the model can distinguish their intrinsic structural responses from the interaction between them. Thus, relevance is determined not by the sheer number of slabs, adsorbates, or reaction pathways included, but by whether the data set contains the right classes of configurations for the scientific question being asked. The third question is coverage. Even when the data set is relevant, the model may still fail if the sampled domain is too narrow. For a narrowly defined problem, the required coverage may remain highly system-specific. For broader and more transferable models, however, the challenge becomes much greater: one must cover not only the target system itself, but also enough neighboring chemistry and configurational variation to avoid brittle generalization. Figure summarizes the idea of relevance and coverage visually: the best-performing model is obtained not from the largest data set, but from a data set that is both sufficiently relevant and sufficiently broad within that relevant domain. In other words, a well-trained model requires not only the right data, but enough of the right data. Schematic illustration of the roles of relevance and coverage in data set design. The most useful training set is not simply the largest one, but the one that remains focused on the scientific scope of interest while sampling that relevant domain broadly enough to support reliable prediction. The green line, blue line, and shaded region represent target, model prediction, and uncertainty, respectively. By “smarter data”, we mean data sets whose composition is deliberately designed so that scope defines the question, relevance defines what matters, and coverage defines how much of that domain must be sampled. Scope without sufficient relevance leads to poorly targeted data generation; relevance without sufficient coverage produces brittle task-specific models; and coverage without relevance produces broad but unfocused data sets. With this lens in place, current data generation strategies can be understood less as isolated workflows and more as different ways of navigating the same central challenge: how to build data sets that are realistic enough for catalysis, efficient enough for practice, and broad enough to support the next generation of catalytic MLPs. This framework is intended as a decision lens for sampling-centered data set design rather than as a universal numerical metric. It converts an open-ended data-generation problem into a constrained sampling problem: scope defines the target scientific domain, relevance identifies the classes of configurations that must be represented, and coverage specifies how broadly those configurations should be sampled. Distribution shift, data efficiency, and uncertainty also provide important statistical diagnostics for sampling-centered data set design, but in this perspective, we will be focused more on sampling than statistical evaluations. In the next section, we discuss why achieving this balance is particularly difficult in heterogeneous catalysis. The Challenges of Data Generation for MLPs in Heterogeneous Catalysis Reaction lies at the heart of heterogeneous catalysis; yet making an MLP genuinely reactive remains one of the hardest challenges in the field. Modern MLPs are anchored in the locality hypothesis: the total energy is decomposed into a sum of atomic contributions, each depending only on a finite radius local environment around an atom, shown in Figure . This “local atomic environment” view supplies a practical route to scalability, linear in system size with short-range cutoffs. In heterogeneous catalysis, however, this same locality-based representation interacts in a nontrivial way with the slab models that underlie most first-principles descriptions of catalytic surfaces. In a typical four-layer 4 × 4 slab may contains dozens of slab atoms, whereas the reacting adsorbates and nearby active-sites may involve only a handful of atoms. As a result, most local environments in the training data correspond to relatively common slab-like configurations than to the sparse but chemically decisive reactive environments. This imbalance matters because many MLPs are trained using atom-averaged objectives. , Under such training, optimization can become biased toward the much more numerous slab environments than the rare reactive environments that control barriers and mechanisms. Thus, if the model is expected to describe reactions rather than only equilibrium structures, reactive configurations must usually be sampled explicitly. − These include transition states and high-energy nonequilibrium structures that are thermodynamically rare but essential. Reactive MLP construction therefore typically relies on targeted, model-aware sampling strategies guided by chemical intuition about the relevant reaction space, which is one reason why reaction exploration remains among the most difficult aspects of MLP development in heterogeneous catalysis. How MLP learns from DFT reference data. (a) Representative atomic structure used to compute DFT ground-truth as inputs. (b) For each atom, the local environment is defined by neighboring atoms within a finite cutoff radius Rc . (c) These local environments are encoded into symmetry-preserving descriptors and fed into a neural network. The network predicts atomic energies (which sum to the total potential energy) and is trained by minimizing the loss against the DFT labels (forces, energies, and stresses). A second reason why data generation remains especially challenging in heterogeneous catalysis is that, unlike many other AI-enabled fields, it still lacks a broadly reusable data set ecosystem despite its central importance. Materials science has developed increasingly rich public databases. Crystalline materials can draw on resources such as the Materials Project, OQMD, and ICSD, while molecular machine learning has been accelerated by data sets such as QM9, and ANI. Heterogeneous catalysis, by contrast, still lacks an equally mature data infrastructure that can be reused broadly across systems. This is not simply because the field has invested less effort, progress such as the Open Catalyst Project exist, but because the underlying combinatorial burden is intrinsically more severe. Even when the catalyst composition is fixed, the relevant chemical space already expands across facets, defects, adsorbates, coverages, reconstruction, and environmental conditions. Under realistic operation, moreover, the catalyst is not a single structure but an evolving ensemble. As a result, the target configuration space is not only vast, but fundamentally open-ended, which makes data set construction in heterogeneous catalysis qualitatively different from the more standardized settings of molecules or bulk materials. Taken together, these two difficulties explain why catalytic MLP building is rarely organized around broad chemical completeness and instead remains dominated by in-house data set construction. Because the decisive reactive configurations are hard to sample and task-dependent, while the overall structural space is too open-ended to be covered by a broadly reusable public data foundation, data set generation is usually carried out backward from a predefined scientific question rather than forward from a universal notion of coverage. This is why current practice in heterogeneous catalysis is still dominated by structure-based, problem-driven workflows: one first narrows the scope to a specific catalyst system or scientific question and only then expands sampling within that relevant domain. Current Practice: Structure-Based Sampling Once the scope is restricted to a specific catalyst system or scientific question, building a relevant data set becomes substantially more tractable. In practice, one typically fixes a catalyst material, a limited set of surface motifs, an operating regime, or a targeted reaction network, and then constructs a seed data set specifically for that intended deployment domain. Such a seed set is often assembled from two complementary sources: publicly available data or pretrained priors, which provide useful structural and chemical starting points, and in-house model construction, where physically motivated slabs, defects, interfaces, adsorbates, and reactive motifs are built explicitly for the catalyst system of interest. In this sense, relevance is rarely inherited directly from public data sets alone; rather, it is established by selecting, adapting, and augmenting available data so that the training set reflects the specific chemistry the model is meant to describe. At the same time, however, a relevant seed data set is rarely sufficient on its own. Even when the initial structures are well chosen, the model may still inadequately cover thermally distorted environments, metastable minima, reconstruction pathways, rare adsorbate arrangements, or reactive configurations that lie just beyond the hand-selected starting library. The common practice in heterogeneous catalysis is not simply to assemble relevant structures, but to begin from a relevance-first seed set and then expand its coverage in a controlled and problem-aware manner. Targeted expansion methods are best classified not by their algorithmic details, but by the type of coverage deficiency they are intended to repair. Perturbation-based methods are most useful when the seed data set already contains the correct structural motifs but lacks sufficient sampling of nearby thermal fluctuations, alternative local minima, or higher-energy regions of the potential-energy surface (PES). Generative methods, by contrast, are more useful when the initial library itself is too narrow and qualitatively new candidates must be proposed, such as new defect arrangements, surface reconstructions, or competing phases. The practical choice of method is therefore less a matter of algorithmic preference than of what kind of missing coverage one is trying to recover within the catalytic domain of interest. Among perturbation-based approaches (Figure ), molecular dynamics (MD) is most effective for enriching coverage around thermally accessible local environments, finite-temperature disorder, and early stage bond fluctuations near known structures. MD is often paired with enhanced sampling (metadynamics, , umbrella sampling) to increase coverage of high-energy regions of the PES. Basin hopping is useful when one wishes to move beyond the local harmonic basin sampled by MD and more efficiently catalog nearby metastable minima. Stochastic surface walking (SSW) is particularly well suited for exploring reconstruction pathways, transition-state candidates, and other directionally accessible regions of the PES. Among generative approaches, genetic algorithms (GA) are especially valuable when broader coverage is needed over qualitatively distinct structures, compositions, defect patterns, or competing phases that are difficult to reach through local perturbation alone. Across all these methods, the underlying principle is the same: they are not used simply to make the data set larger, but to patch specific coverage gaps while remaining anchored to the original catalytic question. Graphical illustrations of approaches to increase data diversity. (a) MD simulates atomic motion under certain temperature and pressure. (b) Basin Hopping “blows open” structures with energetic kicks and then quenching to the nearest minimum. (c) SSW method. Every single step utilizes the climbing module to move uphill and the relaxation module to locate minimum. The green, orange and yellow curves represent the real 1-D PES, the Gaussian functions and the searching trajectory, respectively. (d) Examples of crossover and mutation in genetic algorithms. Parent structure A for crossovers and mutations; parent structure B for crossovers. In practice, these generators are often most effective when paired with active learning. Rather than serving as a separate exploration method, active learning acts as an adaptive selection-and-labeling layer on top of MD, SSW, basin hopping, GA, and related workflows, promoting only those configurations that are uncertain, novel, or insufficiently represented for expensive first-principles labeling. Active learning enhances coverage by steering exploration toward under-sampled regions of the relevant domain, while also preserving relevance by keeping the workflow tied to the catalyst, adsorbate, and operating regime of interest. Representative concurrent-learning frameworks such as DP-GEN make this logic explicit through iterative cycles of training, exploration, selection, and labeling, while related platforms such as GDPy extend similar automation to heterogeneous-catalysis-oriented systems, including surface, metal oxides, supported clusters, and solid–liquid interfaces. More broadly, the shared lesson across these workflows is that active learning does not simply increase data set size; it helps decide which newly generated structures are actually worth paying to label. Common selection signals include committee disagreement among independently trained potentials, force or energy deviation thresholds, descriptor-space novelty, and extrapolation indicators. In heterogeneous catalysis, stopping criteria are necessarily task-dependent: an active-learning loop may be considered sufficiently converged only when model uncertainty remains low across the intended MD, structural search, or reaction-search trajectories, when newly selected configurations become rare, and when downstream quantities such as stability trends, reconstruction pathways, or reaction barriers become insensitive to further labeling. These criteria are more difficult to define in catalysis than in simpler domains because the explored configurations may move across structural, coverage, interfacial, and reactive regimes. These workflows define the current practice of catalytic MLP construction: begin from a seed structural data set, expand coverage with problem-aware exploration, and use active learning to keep labeling focused on the most informative configurations. This strategy remains highly effective because it aligns data generation directly with the scientific question of interest, allowing accurate models to be built for defined catalyst systems and reaction environments. Case Studies There is no universal recipe for constructing a tailored training data set for reactive heterogeneous catalysis. Rather, recent work has produced a series of successful custom-data set workflows using different sampling strategies. Viewed through the present scope–relevance–coverage framework, these studies are especially instructive because they make explicit how data generation choices are shaped by the underlying scientific question. Below, we briefly highlight representative examples that embody this problem-driven philosophy. Lian et al. exemplify this logic in a phase-evolution problem. Their scope was to quantify how much oxygen persists in oxide-derived copper (OD-Cu) under CO2-reduction-relevant conditions, where it resides, and how pH, potential, and specific surface area govern its thermodynamics and kinetics. Relevance was therefore defined by the structural corridor connecting Cu, Cu2O, and the intermediate CuxO states. Data coverage was built in two active-learning stages: independent active learning loops for the Cu and Cu2O endmembers (Figure ), followed by targeted sampling of intermediate CuxO structures. Across 433 active learning iterations, 8.30 × 107 candidates were screened by a two-NN committee, yielding 59,491 labeled points with ∼4.6 meV atom–1 energy and ∼63 meV Å–1 force errors. Embedding pH and electrode potential via the computational hydrogen electrode (CHE) and scanning particle size, the model produced oxygen-depth profiles and desorption-energy statistics that explain residual-oxygen persistence and its tunability, linking the phase-segmented → recombined active learning plan directly to the mechanistic question. Active-learning workflow used to build the training data set for Cu/Cu2O developed by Lian et al. Seed structures (bulk, surfaces, interfaces) initiate iterative cycles of first-principles labeling, model training, and ML-driven exploration. An uncertainty/extrapolation test (e.g., committee disagreement or force/energy thresholds) promotes only the most informative configurations for relabeling. Reproduced from ref. Copyright 2024, Springer Nature. Xu et al. instead defined the scope around support-mediated sintering and dynamic metal–support interactions (MSI) in Cu nanoparticles on γ-/α-Al2O3. Here, relevance was determined by supported thermally accessible structures, Cu-only structures, Al2O3-related surfaces, and supported Cu/Al2O3. Coverage was achieved through an active-learning workflow that combined a genetic-algorithm global search for low-energy supported configurations with finite-temperature MD on the two alumina polymorphs, while model deviation and geometric diversity were used to keep the data set compact but representative (Figure ). The final data set contained 147,464 structures and reached ∼0.004 eV atom–1 (energy) and ∼0.057 eV Å–1 (forces), enabling long-time scale sintering simulations. With this training set, the study resolved termination-dependent compactness across γ-Al2O3 (100)/(110) and α-Al2O3 (0001) and identified a dynamic MSI mode on α-Al2O3 (0001) in which surface Al atoms transiently coordinate moving Cu, outcomes enabled by the targeted sampling of supported, thermally accessible configurations. An active learning pipeline utilizing a genetic algorithm and AIMD for building a data set involving nanoparticles, bulk, surface, and supported nanoparticles focusing on the metal–support interaction. Reproduced from ref . Copyright 2025, American Chemical Society. Zhou et al. considered a broader operando scope: the intrinsic activity of small Au nanoparticles supported on ceria under a CO-oxidation atmosphere, with particular emphasis on size effects, gas-induced disorder, and mechanism switching. This scope required a richer definition of relevance than either of the previous two cases. The data set had to include not only catalyst structures, but also Au/CeO2, clean Au facets, CO and O2 in the gas phase, CO adsorption on facets/clusters/NPs, Au/CeO2 with varied vacancy densities, and single-atom Au sites. Coverage was built from more than 120 AIMD trajectories spanning exactly these environments; frames were down-sampled every ten steps to reduce redundancy, and PCA of learned embeddings was used to quantify diversity in local-environment space (Figure ). The trained potential achieved ∼1.85 meV atom–1 energy error and <89 meV Å–1 force error before being used for 300–1150 K MD across a size ladder of supported NPs. Because the data set was explicitly multiadsorbate (CO/O2/CO2) and vacancy-aware, the model enabled the following size-dependent predictions: ∼3 nm Au remains fcc at 675 K with or without CO, whereas ∼2.3 nm Au disorders markedly on CO exposure, shifting activity from an interface-bound Mars-van Krevelen (MvK) picture toward an associative Langmuir–Hinshelwood route and rationalizing the observed size-phase-activity trend. PCA embedding of structural fingerprints from the training set for CO oxidation on Au/CeO2 catalysts. The data set spans diverse Au facets and Au-CeO2 interfaces and includes multiple CO coverages to capture molecule-surface interactions. Reproduced from ref . Copyright 2024, American Association for the Advancement of Science. Shi et al. provide a particularly instructive example of reaction-network-focused data set design. Their scope was a coupled problem of active-site definition and pathway selection in methanol synthesis from CO2/CO mixtures on Cu–Zn catalysts. Relevance was therefore determined simultaneously by catalyst state and reaction chemistry: the underlying CuZnCHO potential had to support a global PES spanning Cu/Zn/O/H/C structural environments, while cover the downstream application of thermodynamically favorable Cu(111), Cu(211), and Zn-alloyed Cu(211) surfaces as well as the reactive compositions needed for competing hydrogenation pathways. The final training set was generated by SSW-NN, containing 63,539 structures and achieved errors of 2.99 meV atom–1 in energy and 0.080 eV Å–1 in force. In the deployment stage, this potential enabled a microkinetics-guided machine learning pathway search (MMLPS), which expanded a separate reaction-pair database through massively parallel SSW-RS branches initiated from different reactant compositions and surface states, with pathway ranking performed on-the-fly by microkinetic analysis. The model identified the low-Zn-covered Cu(211) step sites as the active motif, show that higher Zn coverages poison the catalyst, and confirm that CO2 rather than CO hydrogenation dominates methanol formation. Taken together, these studies illustrate a common lesson: the success of current catalytic MLPs comes from making the scope explicit, encoding relevance directly into the seed data set, and then expanding coverage only where the targeted scientific problem requires it. At the same time, they also expose the central limitation of the current paradigm: it is effective precisely because it is system-specific and structure-based, but that efficiency does not scale well. Toward Broader Scope: Interaction-Based Sampling Once the intended scope expands, maintaining both relevance and coverage quickly drives up the sampling cost. Even in the relatively narrow case studies discussed above, the final data sets already range from tens of thousands to more than one hundred thousand labeled structures. This is a manageable cost when the scientific question is tightly defined, but it becomes increasingly difficult to sustain if one seeks broader generality. For example, one may require a model and data set to explore surface reconstructions and morphologies across facets and defects, and a separate, highly system-specific sampling strategy to resolve reaction pathways under reaction conditions. Attempting to fold all these complexities, including surface phase space, coadsorbate interactions, and varying solvent/electrolyte environments, into a single, unified MLP remains very challenging. This pressure motivates interest in data generation strategies that can broaden support more efficiently than purely system-by-system, structure-by-structure construction. Recent works have highlighted newly emerging interaction-based sampling strategies, where the objective is not to enumerate all relevant global structures, but to broaden support more directly in atomic interaction space. The basic idea is closely aligned with how modern MLPs learn. Because these models operate on finite-cutoff local environments, broader transferability depends not only on having seen more structural configurations, but also on having seen a wider and more representative range of chemically meaningful local interactions. In the molecular domain, the ANI-1xnr nanoreactor workflow used stochastic high-temperature MD together with active learning to provoke bond making and breaking in random C/H/N/O mixtures, thereby populating reactive local environments without relying on hand-curated mechanisms. In crystalline materials, Liu and coworkers adapted random structure search workflows based on AIRSS , to generate broad structural diversity under light physical constraints, teaching the model both low-energy and unfavorable local coordination across systems ranging from elemental Si to TiO compounds. Although these approaches differ in implementation, they share a common principle: randomness is introduced not to sample one predefined catalytic system exhaustively, but to enlarge support in the underlying interaction space. A recent catalysis-specific example of this broader interaction-based sampling philosophy is the REICO-EMLP framework (Figure ). Rather than beginning from a fixed library of slabs, adsorbates, intermediates, or reaction pathways, REICO aims to generate training data more directly in elemental space. Small cells are initialized with randomized compositions and atomic arrangements, subject only to minimal physical constraints such as interatomic distances, so that the resulting data set can expose the model to a broad range of interelement interactions. This motivation is particularly compelling as configurations that control chemistry, bond-breaking and bond-forming events, transition-state-like environments, and other high-energy nonequilibrium data, are hard to sample, as we noted above. Yet these same configurations disproportionately determine the quantities that matter most in catalysis, including barriers, selectivity, and mechanistic insights. REICO is relevant in this context because it targets a sampling bottleneck that is difficult to address with purely system-specific workflows: sparse but chemically decisive reactive local environments under broader chemical scope. It is also worth noting that, at present, interaction-based sampling should be viewed less as an established replacement for structure-based workflows than as an emerging direction whose eventual utility in heterogeneous catalysis remains to be tested more broadly. (a) Workflow of training element-based MLP (EMLP) powered by the REICO sampling method that focuses on atomic interactions. (b) Example structures generated by REICO method. (c) The resulting EMLP can be used for on-the-fly calculations of structures not directly targeted during training and was evaluated in several heterogeneous-catalysis-relevant benchmark tasks. Reproduced from ref. Copyright 2025, Springer Nature. At the conceptual level, shown in Figure , structure-based and interaction-based sampling should not be viewed as directly rankable alternatives because they do not sample the same object. Structure-based sampling is typically more directed and system-anchored: it follows a predefined catalytic problem and builds relevance directly into the data set by sampling catalyst-specific configurations. By contrast, interaction-based sampling is more distributed and space-covering: it samples local reactive or elemental environments across a broader chemical domain rather than along a single system-specific trajectory. An apples-to-apples comparison between the two is therefore not well-defined. Broader local-environment coverage does not automatically guarantee relevance at the system level: catalytic behavior can still depend on surface identity, correlated coverage patterns, collective reconstruction, support interactions, and environmental context that are not fully specified by local environments alone. A practical hybrid workflow could use interaction-based sampling as one source of broad local-interaction coverage, while retaining system-relevant structures as anchors for catalyst-specific relevance. Similar caution applies to systems involving long-range electrostatics, charged interfaces, or other effects that are not easily reduced to short-range tiling. For this reason, the most realistic role of interaction-based sampling is likely within a hybrid workflow. Structure-based sampling remains strongest when the scientific question is sharply defined, because it encodes relevance directly into the data set. Interaction-based sampling becomes attractive when broader scope and transferability are desired, because it can improve coverage and balance in reactive local-environment space while reducing dependence on fully tailored data sets for every nearby system. Conceptual illustration of structure-based and interaction-based sampling. Colored curves denote the projected chemical-space trajectories of different systems. Structure-based sampling follows predefined system-specific paths and therefore builds system-level relevance directly into the data set, but requires separate sampling for each system. Interaction-based sampling instead seeks broader transferable coverage across systems through shared underlying physical interactions, rather than being tied to a single trajectory, but it does not inherently guarantee the same degree of system-specific relevance. Conclusion and Outlook Heterogeneous catalysis inherently couples multiple layers of complexity, including structural reconstruction and surface coverage effects. This creates an intrinsic mismatch between the complexity of the problem and the limitations of conventional theoretical tools: DFT is costly, whereas simplified models often cannot capture the full complexity. As a result, heterogeneous catalysis remains one of the most demanding application domains for MLPs. However, the reliability of MLPs depends not simply on model architecture, but on whether the training data distribution is relevant to the scientific question, sufficiently broad to cover the intended chemical space, and rich enough to represent the rare but decisive environments that control catalytic behavior. From this perspective, the development of MLPs in heterogeneous catalysis can be read as an evolution in data generation: from borrowing from public databases, to expanding structure libraries, to constructing tightly targeted system-specific data sets, and more recently to exploring broader interaction-centered strategies. The common lesson is that progress will not come from ever-larger data sets alone, but from smarter data. At the same time, no single data generation philosophy is likely to solve the problem on its own. Structure-based workflows remain indispensable because they encode relevance directly into the data set and still provide the most reliable route to system-specific accuracy. Interaction-based sampling becomes attractive when the intended scope expands and purely system-based workflows begin to strain, especially when broad reactive local-environment coverage and transferability across related configurations are required. The most realistic path forward is therefore likely hybrid: broader reactive support combined with catalyst-specific anchoring, rather than a simple replacement of one approach by another. As this perspective has focused primarily on the sampling range of catalytic data sets, several additional challenges remain beyond the present scope. One is data labeling, which ultimately determines the fidelity of the trained MLP. In particular, it remains important to understand how the choice and consistency of electronic-structure labeling affect MLP reliability. , Multifidelity workflows are particularly interesting as they combine broad low-cost exploration with selective high-fidelity relabeling of the most chemically consequential configurations. Such a strategy offers a practical route to preserve coverage while concentrating the highest labeling accuracy on the regions that matter most. Diffusion-based methods have begun to appear in adjacent areas relevant to data set design, especially transition-state generation/search, diffusion-inspired MD and the generation of novel candidates optimized for desired criteria, although direct applications to catalytic data set generation are still limited. Another important direction is how to avoid building every catalytic MLP entirely from scratch. Pretrained and foundation models − offer an attractive starting point, and fine-tuning with limited new data may provide an efficient route toward domain-specific improvement. However, important questions remain open regarding task-specific adaptation, including bias transfer, data-mixing effects, and catastrophic forgetting during refinement. More broadly, new data regimes also raise new requirements for training discipline and evaluation. Recent data sets have increasingly targeted off-equilibrium configurations in order to improve generalization to strained, defective, and thermally distorted environments. Yet such deliberate distribution shifts also challenge existing assumptions about training stability, model robustness, and performance assessment. The field will therefore need better evaluation standards and more operational data set quality metrics for these increasingly diverse training distributions. The central message is that atomistic machine learning is entering a stage where data set design can no longer remain implicit. Heterogeneous catalysis makes this especially clear because it exposes, in an unusually sharp form, limited reusable data, sparse but chemically decisive reactive structures, narrow training domains, and an open-ended configurational space. Together, these features make it difficult to maintain relevance and sufficient coverage at the same time, especially as scope grows. The future of MLPs in heterogeneous catalysis may therefore come not only from building more powerful models, but by how well we can decide what data are actually worth generating.