WorksPhysics and Astronomy
A Critical Examination of Active Learning Workflows in Materials Science
Abstract
Active learning (AL) plays a critical role in materials science, enabling applications such as the construction of machine-learning interatomic potentials for atomistic simulations and the operation of self-driving laboratories. Despite its widespread use, the reliability and effectiveness of AL workflows depend on implicit design assumptions that are rarely examined systematically. Here, we critically assess AL workflows deployed in materials science and investigate how key design choices, such as surrogate models, sampling strategies, uncertainty quantification and evaluation metrics, relate to their performance. By identifying common pitfalls and discussing practical mitigation strategies, we provide guidance to practitioners for the efficient design, assessment, and interpretation of AL workflows in materials science.
Introduction
The application of machine learning (ML) in materials science and engineering has expanded rapidly, enabling progress across a wide range of tasks, including high-throughput screening, accelerated materials property prediction and autonomous experimentation [1, 2, 3, 4]. Central to many of these efforts is the challenge of learning complex mappings between structure, composition, processing conditions, and materials properties, which are typically non-linear and only partially understood. By learning directly from data rather than relying on explicit physical models, ML methods provide a flexible framework for capturing such complex relationships and have demonstrated transformative potential in materials science [5, 6]. However, their success is fundamentally constrained by the availability, accuracy and quality of the training data. Despite advances in high-throughput experimentation and simulation frameworks [7, 8], generating consistent, high-fidelity materials data remains time- and resource-intensive. As a result, many materials-science applications operate in regimes better characterized as data-scarce rather than data-rich [9]. Moreover, materials datasets are often not “statistically representative”, i.e., they tend to be shaped by prior domain knowledge, biased curation strategies, or feasibility constraints, and frequently fail to adequately represent the broader space of relevant structures, compositions, or other relevant parameters. As a result, ML models trained on such data are prone to unreliable performance when applied to unexplored or underrepresented regions of the data space [10, 11].
Active machine learning (hereafter referred to as active learning, AL) has emerged as a promising framework for addressing these challenges in ML-driven materials science. As formalized by Cohn et al. [12], “AL studies the closed-loop phenomenon of a learner selecting actions or making queries that influence what data are added to its training set”. By employing adaptive sampling strategies, AL prioritizes the acquisition of critical data points [13]. AL is often conceptualized as an efficient labeling strategy. Its applications in materials science are broad, ranging from training data generation [14, 15] to guiding closed-loop experimentation and autonomous materials discovery [16, 17]. To illustrate the diversity of AL use cases, we briefly highlight two representative application settings that are currently of significant interest: (i) ML-driven atomistic simulations using machine-learning interatomic potentials (MLIPs) and (ii) optimization-based workflows for materials discovery. In the context of MLIPs, a central challenge is capturing rare but physically important phenomena such as bond breaking, phase transitions, or defect migration that occur infrequently, which are often poorly represented in training datasets. Here, AL can be used to ensure the inclusion of underrepresented yet physically essential configurations, particularly those associated with rare events or extreme conditions encountered during MLIP-guided molecular dynamics (MD) simulations [18, 19, 20]. In autonomous laboratories focused on materials property optimization, the primary difficulty lies in identifying statistically exceptional materials within vast design spaces (discussed in detail later). In this setting, AL aids in achieving the goal of maximizing the likelihood of selecting promising candidates for evaluation while minimizing the cost required to find materials with certain properties. Despite these differing objectives, AL is typically implemented as a workflow composed of intertwined design choices, such as surrogate models, sampling strategies, uncertainty quantification and performance evaluation metrics.
While AL workflows are now routinely applied in materials science [21, 22, 23], their design choices vary widely, even for closely related tasks. This diversity partly reflects the aforementioned breadth of materials science applications and is therefore neither surprising nor inherently problematic. However, it also introduces substantial inconsistency in how AL workflows are designed, executed, and assessed, making it difficult to compare performance or draw general conclusions across studies. For example, in data generation tasks, AL workflows may be designed to prioritize unfamiliar data points (e.g. novel compositions or structures) [24], reduce biases in the initial dataset [25], improve predictive performance of ML models [26], or decrease model uncertainty [18]. While these objectives are often interrelated, practitioners typically evaluate the effectiveness of AL using only a subset of metrics, most commonly improvements in model accuracy, without systematically assessing whether coverage of critical regions of the data space has also improved. Similarly, in materials discovery applications, the choice of surrogate models [10], acquisition functions (vida infra) [27], and uncertainty quantification methods [28] can strongly influence outcomes, yet these choices are often made heuristically and evaluated using application-specific criteria that hinder cross-study comparison. Moreover, a fundamental question remains insufficiently addressed: “to what extent does AL provide benefits beyond those achievable through simpler data selection strategies based on human intuition, prior domain knowledge, or hand-crafted rules?.” In other words, it is often unclear whether observed performance gains stem from a principled, algorithmic AL framework or could instead be achieved by informed, human-guided selection of data points without an explicit AL loop. These issues highlight the need for modular and standardized AL workflows that enable consistent evaluation and comparison across applications, while still allowing flexibility to accommodate domain-specific objectives. Without such structure, AL workflows risk being inefficient, difficult to interpret, or misdirected, potentially leading to unnecessary computational or experimental cost and convergence toward suboptimal solutions.
In this perspective, we critically examine AL workflows commonly employed in materials science, focusing on two key applications of data generation and materials discovery. We begin with a brief overview of the AL methodology and analyze the strengths and limitations of tools and techniques used at different stages of AL workflows. We do not aim to provide a comprehensive review here and instead refer readers to existing review articles for the broader context [21, 29]. Our focus is to raise awareness of practitioner-driven biases in the design choices that can impact the performance of AL workflows. Building on this analysis, we propose guidelines to support the rigorous design, assessment, and interpretation of AL workflows in materials science.
Active Learning Methodology
An AL algorithm involves performing data generation iteratively and adaptively, with the goal of selecting the most important data points for labelling under a limited evaluation budget. In materials science, the “labels” are typically obtained from an oracle such as a high-fidelity simulation, a physical experiment, or expert annotation (Figure 1a), all of which are costly in terms of time or resources. An AL workflow therefore seeks to maximize learning efficiency by prioritizing which data points to evaluate next, rather than relying on sampling performed randomly [30]. Because AL algorithms often construct a sequence of targeted queries, it is sometimes referred to in the literature as “query learning” [31, 32] or “sequential learning” [33, 34]. Several AL variants exist depending on how unlabeled data are accessed or generated, including query synthesis, where new candidate inputs are generated by the learner [35], and stream-based AL, where data arrive sequentially and must be queried or discarded on the fly [36]. However, we will use the term “active learning” and stick to discussing pool-based AL [37] where it is assumed that a relatively large pool of unlabelled data is available for labelling, because of its simplicity and its widely adopted usage in the field. The objective of an AL algorithm is to achieve a task-specific goal (e.g. improve model performance) while minimizing the overall cost of labeling (or another user-defined criterion) through strategic sampling (Algorithm 1). In a supervised learning setting, let denote the initially available labelled dataset (hereafter referred to as “seed data”), and let represent the pool of unlabeled data (hereafter referred to as “pool data”). Here, represents the input space (e.g. crystal structures, compositions) and represents the corresponding outputs of interest (e.g. energies, forces, or materials properties). A surrogate ML model is trained to approximate an unknown target function , mapping to . At each iteration, a sampling strategy evaluates the pool data using the current model and selects one or more candidate inputs , deemed most informative according to a predefined criterion. These candidates are then evaluated by an oracle to obtain labels , after which the labeled dataset and unlabeled pool are updated:

Figure 1. Schematic representation of (a) an AL workflow where both model-based and model-free AL strategies can be employed to acquire samples from pool data and update the seed data by interacting with an oracle, (b) various factors that need to be considered while designing the sampling strategy for AL.
This process is repeated until a stopping criterion is satisfied, such as a performance threshold or exhaustion of a query budget .
Algorithm 1 Pool-Based Active Learning Algorithm
Require: Initial labeled dataset $D_{\mathrm{init}} \subset X \times Y$, unlabeled pool $U \subset X$, oracle $\mathcal{O}$, sampling strategy $Q$, budget $B \in \mathbb{N}$, surrogate model $M$
Ensure: Trained model $M$ approximating $f:X \to Y$
1: Train surrogate model $M$ on $D_{\mathrm{init}}$
2: while stopping criterion not met and $B > 0$ do
3: Select query point $x^* \leftarrow Q(M,U)$
4: Obtain label $y^* \leftarrow \mathcal{O}(x^*)$
5: Update dataset $D_{\mathrm{init}} \leftarrow D_{\mathrm{init}} \cup \{(x^*,y^*)\}$
6: Remove queried sample $U \leftarrow U \setminus \{x^*\}$
7: Retrain $M$ on updated $D_{\mathrm{init}}$
8: $B \leftarrow B - 1$
9: end while
10: return Trained model $M$
Since the AL algorithm involves weighing of unlabeled data and strategic selection of the best samples for labelling, it offers a targeted approach that allows the model to learn efficiently from fewer samples, which is critical for a data-limited domain like materials science. More generally, AL provides a data-efficient framework for exploring high-dimensional design spaces, i.e., the set of candidate inputs under consideration for a given task (e.g. possible alloy compositions). By integrating model training and data generation within a single closed-loop workflow, AL supports targeted exploration of these spaces without resorting to exhaustive enumeration, thereby enabling the identification of promising materials or configurations at reduced computational or experimental cost. The iterative nature of AL further allows models to adapt as new data are acquired, reducing reliance on initial, potentially biased datasets as learning progresses.
Active Learning Applications and Challenges
In this section, we highlight the key challenges in employing AL workflows for materials science problems through two representative application domains: (i) materials data generation and (ii) materials discovery. It is to be noted that here, the term “materials discovery” refers broadly to identifying materials with desirable properties, whether among previously synthesized, computationally generated, or yet-to-be-synthesized candidates. In this sense, discovery entails jointly establishing material identity and properties within a vast, largely unexplored materials space. For both i) and ii), we discuss, based on existing literature, how AL can be useful and identify critical challenges that remain unaddressed or are frequently overlooked. We provide practical guidelines and recommendations for addressing these challenges in future AL-driven materials research in the Outlook section.
Data Generation/Redundancy Mitigation
Redundancy problem in materials data: Materials datasets are often curated based on prior knowledge (e.g. well-known materials with desirable properties) or ease of access (e.g. from existing data repositories). These practices can introduce substantial redundancy and lead to biased coverage of the broader materials space. Recent work by Li et al. [38] demonstrated that widely used computational materials databases, including the Materials Project (MP) [39] and OQMD [40], contain significant data redundancy. Using a data-pruning strategy, they showed that removing a large fraction of redundant entries neither degraded in-distribution model performance nor improved out-of-distribution (OOD) generalization for ML models trained on the full datasets (Fig. 2a).
While such redundancy is perhaps unsurprising given the high-throughput nature of computational database generation, biased sampling can persist even when subsets of these databases are selectively curated. For example, in many materials discovery studies [6, 41], candidate materials are preferentially chosen near the convex hull of a given compositional phase diagram, which can distort an ML model’s representation of the underlying stability landscape. Indeed, Bartel et al. [42] showed that ML models can accurately predict formation energies yet still perform poorly in classifying stable versus unstable materials, particularly in sparsely sampled compositional spaces such as underrepresented quaternary systems in MP. In such scenarios, AL offers a promising alternative by adaptively focusing on data in the underexplored or informative regions of the materials space. By doing so, AL can enable the construction of smaller, more relevant datasets that mitigate redundancy while maintaining or even improving predictive performance.
Inadequate sampling strategy: One of the key components of an AL workflow is the sampling strategy which is used to assign the aforementioned relevance of samples from an unlabelled pool. Many of the AL workflows adapted in materials science [26, 43], especially those employed for MLIP training [44, 45] uses informativeness, i.e., ability of a sample to improve the model performance, as the sole sampling criterion. However, an effective AL sampling strategy is inherently multi-faceted and cannot be fully captured by informativeness alone (Fig. 1b). A key complementary criterion is representativeness, which assesses whether a queried sample reflects the structure of the unlabeled data distribution [46]. This is particularly important to prevent sampling extreme outliers (e.g. highly distorted structures) that are not statistically representing the remaining, relevant materials space of interest. Various methods have been proposed by the ML community to include the representativeness factor [47, 48], or jointly optimize informativeness and representativeness [49, 50], yet most AL workflows in materials science remain focused on informativeness due to the ease of monitoring via proxies such as model test errors. Beyond these standard criteria, additional materials-science specific factors need to be considered for sampling. These include: (i) diversity, which ensures that selected samples are sufficiently distinct relative to both the seed data and between themselves to avoid redundancy [51, 52], (ii) physical validity, ensuring that samples are chemically and physically meaningful, for example by excluding unphysical MD configurations or crystal structures under extreme conditions and (iii) feasibility, which accounts for practical constraints such as computational cost, favoring candidates that provide maximal information without incurring excessive expense (e.g. extremely large unit cells). In addition, for AL campaigns for which a reasonably large seed data are already available, pre-existing biases can persist, such as overrepresentation of certain chemistries, phases, or structural motifs. If these biases are not identified and addressed at the outset, AL may inadvertently reinforce them, leading to a sampling that further skews the dataset. Failure to consider these criteria while designing the sampling strategy can therefore severely limit the reliability of AL in materials science.
The ill-addressed active learning bias: While AL can mitigate the redundancy problem, it can paradoxically introduce a new form of bias, termed as active learning bias (ALB) [53, 54]. This arises because during AL, samples are no longer drawn independently and identically distributed (i.i.d.), a fundamental assumption underlying ML model training. As a result, actively curated training sets may deviate substantially from the application-relevant data distribution in the materials space of interest (Figure 2b). This deviation has important implications for model training and evaluation. Standard empirical risk minimization, which optimizes model parameters by minimizing the average loss over the training data, implicitly assumes that the training set is representative of the target distribution. When this assumption is violated, as is often the case in AL, model performance measured on finite or randomly curated test sets may reflect optimization with respect to a biased objective rather than genuine generalization across the materials domain. In materials science applications, this can manifest as models that perform well on actively sampled configurations while failing to generalize to unexplored chemistries or structures. Statistical corrections have been proposed to mitigate ALB, such as reweighting the training loss by the inverse probability of sample acquisition [55], thereby partially restoring consistency with the underlying data distribution. However, such

Figure 2. (a) Active learning reduces redundancy in materials datasets: Performance of XGBoost (XGB) and Random Forests (RF) models on band gap prediction trained on datasets obtained by uncertainty guided active learning, pruning, and random sampling from the OQMD14 dataset. Comparable accuracy is achieved with only 10% of the data, highlighting substantial redundancy in the dataset, Adapted with permission from Ref. [38], Copyright 2023 Springer (b) schematic representation of active learning bias induced by sampling of data points do not following i.i.d assumption, (c) information-entropy guided active learning (ETAL) minimizing the large structure-stability bias by improving the coverage of less symmetric crystal systems in the JARVIS classical force-field inspired descriptors dataset (top panel) and improved performance of ML models trained on such active learned dataset compared to random sampled dataset (bottom panel). Adapted with permission from Ref. [25], Copyright 2023 American Institute of Physics.
approaches have not yet been systematically explored in materials science. Moreover, reweighting introduces additional subtleties: in overparameterized models, including deep neural networks, ALB can sometimes act as an implicit form of regularization, reducing overfitting and even improving apparent generalization [56]. While this effect may be beneficial in practice, it complicates the interpretation of AL performance, as improvements may stem from sampling-induced regularization rather than principled coverage of the materials space. Because the magnitude and impact of ALB depend strongly on the mismatch between the seed dataset, the unlabeled pool, and the target application domain, careful attention to data distributions is essential. Incorporating explicit analysis of distributional coverage using tools such as dimensionality reduction [57] or density estimation [58] can therefore provide critical context for evaluating AL outcomes and for designing more robust, application-aware AL workflows [59, 60]. Without such considerations, AL strategies risk reinforcing hidden biases, limiting transferability, and overstating the effectiveness during the workflow deployment.
Model-free vs. model-based active learning: While most AL workflows employed in materials science are built around surrogate ML models and consequently face the challenges outlined above, model-free strategies could also be adapted for conducting AL as an alternative. A key advantage of model-free strategies is their conceptual simplicity and flexibility; sampling decisions can be designed to target specific objectives without relying on model predictions or uncertainty estimates. This is particularly attractive for scenarios involving a very small amount of seed data, where surrogate models have limited predictive accuracy, and their uncertainty estimates may be unreliable. For example, Information-theory based metrics such as information-entropy and mutual information capture expected information gain beyond model-derived properties [69, 70]. Zhang et al. [25] employed an information-entropy-based AL workflow to mitigate structure–stability bias in computational crystal databases, where low-symmetry structures are often underrepresented. By prioritizing structurally informative samples, measured using information entropy, their approach improved the coverage in crystallographic space and yielded ML models with superior predictive performance compared to random sampling (Figure 2c). Similarly, Schwalbe-Koda et al. [61] demonstrated that atomistic information entropy, computed directly from local atomic descriptors, can serve as a model-free proxy for uncertainty of MLIPs, guiding MD simulations. Beyond informativeness, model-free AL can explicitly enhance representativeness and diversity, two criteria that are often weakly controlled in model-based AL workflows. Density-based strategies, such as clustering or kernel density estimation, promote sampling from statistically significant regions of the materials space, thereby preserving global coverage and facilitating the representative sampling [66, 71]. On the other hand, similarity-based model-free strategies emphasize improving diversity and minimizing redundancy. These methods are often implemented using some distance-based metrics, defined over feature, descriptor, latent, or embedding spaces [38, 72, 73] and select samples that are maximally distinct from one another and from the existing seed data, promoting diversity. Alternatively, diversity can also be enforced using physically or chemically motivated similarity measures, for example, based on composition, local coordination environments, structural topology, bonding motifs, or symmetry classes [68, 73]. It has been shown that when labeled data are scarce, similarity-based model-free methods can outperform model-based AL due to their robustness against inaccurate surrogate models [74]. However, these approaches are not without limitations. High-dimensional vector spaces used to represent materials data may suffer from the curse of dimensionality, and outcomes of distance-based sampling are sensitive to the choice of representations (e.g. elemental-property-based features or SOAP descriptors [75]), similarity metrics (e.g. Euclidean or Mahalanobis distances), and analytical choices such as centroid-based versus nearest-neighbor selection [76]. Recent benchmarking by Bi et al. [77] further suggests that, on average, model-free strategies underperform model-based AL when evaluated across diverse materials datasets, as they lack explicit mechanisms to capture the relationship be-
| Sampling Criterion | Methods | Central Idea | Application in Materials Science |
| Informativeness | Entropy-guided | Select samples that maximize information entropy computed from structural or descriptor distributions, enabling bias reduction and improved coverage without model uncertainty estimates. | Curation of bias-minimized crystal structure datasets [25]; Model-free uncertainty for MLIP-driven MD [61, 62] |
| Mutual-information-guided | Maximize mutual information between candidate samples and the existing dataset or experimental objectives to guide efficient discovery. | Phase-change memory material discovery [63] | |
| Representativeness | Clustering | Partition the unlabeled pool into clusters and select representative samples (e.g. cluster centroids) to preserve global coverage of the distribution and avoid oversampling statistical outliers. | Discovery of perovskite oxides for oxygen evolution catalysis [64]; Training data generation for MLIPs [65] |
| Density estimation | Prioritize samples located in high-density regions of the unlabeled data distribution to ensure representativeness. | Functionalized nanoporous materials (MOFs/COFs) property prediction [66]; Assessing out-of-distribution performance of ML models [58] | |
| Diversity | Distance-based | Select samples that maximize the minimum distance to the seed data in feature, output, latent, or an embedding space to avoid redundancy and maximize diversity. | Discovery of high-entropy oxides for H production [16]; Surface structure exploration for catalysis [67] |
| Physical metric-based | Select batches of samples that are mutually dissimilar while also distinct from the existing labeled data based on a physical or chemically motivated metric | Developing accurate property-prediction models for structure-property mapping of microstructures [68] |
Table 1. Examples of model-free active learning strategies and representative applications in materials science.
tween samples and target properties. Despite these limitations, model-free AL remains practically valuable; it avoids the computational overhead of training and retraining surrogate models (e.g. ensembles of ML models) and remains effective when initial datasets are small or highly biased. These observations highlight an unresolved dilemma in AL design: whether to prioritize model-based or model-free strategies, and motivate hybrid approaches that combine the robustness of model-free sampling with the task-awareness of model-based methods [78, 79]. Representative model-free AL strategies and their applications in materials science are summarized in Table I.
Materials Property Optimization and Discovery
Interplay between surrogate models and sampling strategy: The dominant AL practices in both computational and experimental settings for materials property optimization and discovery are black-box optimization (BBO), with Bayesian Optimization (BO) [87] being the most widely adopted approach [88]. BO’s ability to navigate complex search spaces under data-scarce conditions has led to numerous success stories in materials discovery [89, 90, 91]. Notable examples include the identification of Pb-free BaTiO-based piezoelectrics with enhanced electrostrictive strain [92], NiTi-based shape memory alloys with low thermal hysteresis [84], and efficient high-entropy alloy catalysts [91]. The key components of a BO-driven AL (BO-AL) include: (i) a surrogate model to approximate the expensive objective function mapping the materials property to a set of given input parameters, (ii) an acquisition function (analogous to sampling strategy in non-BO AL) to guide sample selection, and (iii) an oracle for labelling new data points. While BO traditionally relies on probabilistic surrogate models such as Gaussian Processes (GPs) for their principled uncertainty quantification, non-Bayesian models have also been adopted in materials applications [93, 94]. This is often motivated by practical considerations, including the poor scalability of GPs in high-dimensional spaces and the superior extrapolation performance observed with certain non-Bayesian models on specific datasets [95]. Nevertheless, the optimal choice of surrogate model for BO-AL in materials science remains an open question, with conflicting findings reported in the literature. For instance, Lim et al. [93] demonstrated that GPs with carefully selected kernels outperformed alternative models, including random forests (RF), on experimental materials datasets. In contrast, Liang et al. [94] found that RF-based BO outperformed GP-BO using standard isotropic kernels and performed comparably to GPs with anisotropic kernels. Table II summarizes some of the characteristics of various surrogate models used in BO-AL for materials science applications.
| Model Type | Data Efficiency | UQ | Interpretability | Cost | Application in Materials Science |
| Gaussian Processes | High | Principled (exact Bayesian posterior) | Limited in high-dimension | High () | Phase-change memory material for photonic switching devices [63] (exp); layered materials with suitable electronic properties [22] (comp) |
| Random Forests | Moderate | Heuristic (ensemble-based) | Moderate with feature importance | Low | Biochar synthesis for CO capture [80] (exp); screening of inorganic materials [81] (comp) |
| Gradient Boosting Methods | Moderate | Heuristic (ensemble-based) | Moderate with feature importance | Moderate | High-entropy oxides for H production [16] (exp); power factor prediction of thermoelectrics [43] (comp) |
| Bayesian Neural Networks | Low | Heuristic (approximate posterior) | Low | High | Optimal parameters for chemical reactions [82] (exp); van der Waals heterostructures with suitable bandgaps [83] (comp) |
| Support Vector Regression | Moderate | Limited | Limited | Moderate | Shape memory alloys with low thermal hysteresis [84] (exp); Piezoelectric materials screening [28] (comp) |
| Deep ensembles | Low | Heuristic (inter-model predictive variance) | Low | High | Crystal structure prediction [85] (comp); MLIP assisted material simulations [45] (comp) |
| Symbolic regression | High | Limited | High (analytical equations) | High | Screening of acid-stable oxides for electrocatalysis [86] (comp) |
Table II. Common surrogate models and their properties relevant to Bayesian Optimization based active learning applications. The exp and comp in parenthesis of literature references indicates whether the works do involve experiments or purely computational simulations, respectively.
Importantly, the performance of BO-AL is often not governed just by the surrogate model or acquisition function in isolation, but by their combined behaviour. A highly accurate surrogate model may still perform poorly if paired with a suboptimal sampling strategy. For example, expected improvement (EI), a popular acquisition function used in BO-AL for material property optimization, depends on the current best observation, which might be an unreliable benchmark if the seed data is strongly biased, misleading the optimization trajectory. Therefore, evaluating surrogate model–acquisition function combinations through after-the-fact (AFT) AL trials (where pool data is already labelled but excluded from seed data) can help identify optimal configurations [84], though these analyses may not always generalize to all pool datasets due to distribution shifts. As illustrated by Boley et al. [10], in the AFT-AL experiment for discovering perovskites with high bulk modulus, RF with both pure exploitation (XT) and EI acquisition functions perform similarly to that of GP with EI, but only as good as RS. However, in real AL runs (using DFT calculations as ground truth), GP with EI clearly outperforms the others, highlighting that surrogate model–acquisition function selection based solely on initial data may be misleading due to distributional shifts and the underrepresentation of high-performing materials. Although dynamic switching between acquisition functions has been proposed [96], it remains underexplored in BO-AL workflows applied for materials discovery. It is to be noted that such a sensitivity to sampling strategy is a general challenge in AL, even beyond BO frameworks. Liu et al. [68] demonstrated this by analyzing how representation and sampling methods affect the selection of microstructures for property prediction via AL. They found that a sampling strategy considering both input features and model outputs outperformed those based on either alone, including uncertainty-based approaches, highlighting the crucial role of the sampling strategy in determining the AL efficiency.
For many applications, BO-AL is preferred to be applied in batched or parallel settings where a number of samples (instead of one as in sequential BO) are selected, which could leverage the availability of oracles with parallel execution capabilities (e.g. high performance computing facilities, high-throughput synthesis platforms) [23, 82]. While batching is desirable from a resource management perspective, it can risk degrading sample diversity and informativeness unless the sampling strategy is designed to account for these factors [99]. Acquisition functions often fail to correctly rank batched instances in terms of marginal utility, potentially introducing redundancy or bias. This is particularly problematic in materials science, where multiscale material behaviours and divergent synthesis/characterization routes can make it difficult to evaluate a batch under uniform experimental conditions. Nevertheless, platforms such as Pheonix [82], ChemOS [100], Olympous [101] etc. have made success in enabling the application of BO-AL with robotic experimentation or lab automation systems to enable parallel acquisition and execution, useful for chemistry and materials science.
Unreliable uncertainty quantification: In AL strategies with an exploration component such as those used in BO, uncertainty quantification (UQ) plays a pivotal role. This is based on the notion that the surrogate models are most error-prone in regions where their predictions are least confident, and hence, acquiring data from such regions is often the most informative. Reliable UQ helps in identifying these regions by assigning a confidence interval to predictions, thus guiding the sampling strategy toward high-impact data points. However, two core challenges hinder this process: (i) most ML models lack inherent UQ capabilities (see Table II), and (ii) many UQ techniques produce unreliable estimates which are often underconfident or overconfident relative to actual prediction errors (Figure 3b) [102, 103]. Unreliable UQ can misguide AL by undervaluing informative samples, potentially leading to premature convergence or excessive exploitation of suboptimal regions. For instance, in BO-AL with EI as the acquisition function, overconfident uncertainty estimates shrink the exploration term, causing the algorithm to overlook uncertain but informative regions in favor of already well-explored ones.

Figure 3. (a) Impact of surrogate model and acquisition function choices on active learning for identifying perovskites with high bulk modulus: Real AL runs (top panel) show varied efficiencies across GP and RF models with exploitation (XT), expected improvement (EI), and uniform (random) sampling acquisition functions, compared to retrospective seed-only baselines (bottom panel). Adapted with permission from Ref. [10], Copyright 2024 Institute of Physics. (b) Unreliable uncertainty estimates: deviation of z-score distributions (residuals scaled by predicted uncertainty) from the unit normal indicates overconfidence (top panel). In AL-guided computational search of acid-stable oxides with the promising figure of merit (Pourbaix decomposition free energy eV/atom), such overconfident estimates are frequent, despite the success of AL in identifying 12 acid-stable oxides within 30 iterations guided by SISSO [97] as surrogate model and probability of feasibility (POF) as acquisition function (bottom panel). Reprinted with permission from Ref. [86], Copyright 2025 Springer, (c) Lack of correlation between model performance and materials discovery in Bayesian optimization based AL for identifying first decile materials with near-zero band gaps in the dataset. Colored curves denote acquisition functions (blue: expected value, violet: EI, red: maximum uncertainty, yellow: random sampling). Discovery yield (DY) (see Table III for definition) improves, but non-dimensional model error (NDME = RMSE/standard deviation of holdout set) does not consistently decrease. Adapted with permission from Ref. [98], Copyright 2023 Royal Society of Chemistry
The reliability of UQ in ML remains an open research challenge in materials science, with relatively few studies that critically benchmark UQ methods [106, 107]. Existing efforts typically evaluate the quality of uncertainty estimates based on hold-out test sets using metrics based on the uncertainty distribution [108], correlation with prediction errors [106], or computational cost [107]. These studies consistently show that no single UQ method outperforms across all scenarios, with strong dependencies on surrogate model and dataset characteristics. Nevertheless, calibration approaches such as scaling uncertainty estimates with respect to residuals can enhance their reliability [108, 109]. Openly available tools like uncertainty-toolbox [110] and UQLab [111] facilitate such calibration processes. However, the effectiveness of these calibration methods within AL workflows remains largely unexplored, with existing studies still in their early stages [112]. Since such approaches require labeled data, they are constrained to the currently known samples and hence cannot guarantee improved uncertainty estimates on the unlabeled pool data where accurate UQ is most critical. This highlights the need for more standardized, AL-focused metrics to assess and compare UQ methods for materials property prediction and optimization tasks.
Since one of the reasons for unreliable UQ is that the surrogate models are primarily trained to minimize prediction error without providing default uncertainty estimates, a range of alternative, model-agnostic UQ strategies have been proposed to address this challenge. These rely on heuristic measures of uncertainty, such as distance in feature [113] or latent spaces [72] and are not specifically dependent on predictive model architecture. It has been demonstrated that latent space distance provides more reliable uncertainty estimates (by better capturing of residuals) for both artificial [72] and graph [114] neural networks, indicating that such model-agnostic methods could be beneficial to AL. Similarly, integrating Domain of Applicability (DoA) analysis [57, 58] with UQ in AL workflows allows for the identification and expansion of the model’s reliable prediction regime, enabling more effective sample selection near or beyond the current DoA to improve model robustness [115]. However, these methods are more mature in cheminformatics, and lack standardized implementation in materials science, due to the absence of universally accepted DoA frameworks [58, 116].
Also, a fair comparison between these model-agnostic vs. model-based UQ methods for AL performance for materials property optimization has not been done to the best of our knowledge, which is also necessary for improving the reliability of AL practices in materials science.
Lack of standardized performance evaluation metrics: A rigorous and accurate evaluation of AL performance is essential for assessing its efficiency and determining when to terminate the AL loop. This is especially critical in closed-loop, automated experimentation platforms with limited human intervention, where failure to recognize diminishing returns can result in substantial resource waste. Various performance metrics have been proposed to quantify AL efficiency in materials property optimization. However, no single metric universally captures all aspects of performance, and even metrics which are expected to be complementary could show diverging trends for performance estimation throughout AL iterations. The most popular metrics used in assessing the efficacy of AL are based on comparison with random sampling (RS) [93, 104], which involves sampling data points from the pool without the use of a surrogate model derived insights and hence also can be referred to as “passive learning”. Rohr et al. [33] proposed quantitative metrics such as the acceleration and enhancement factors (see Table III), measuring the fraction of promising materials found and the iterations required to find them relative to RS, which have been used to benchmark surrogate models and acquisition strategies for AFT materials discovery campaigns [94, 117]. However, using RS as a baseline may not be considered a universal performance metric in materials science, where experimental choices are typically guided by prior knowledge rather than random selection. Moreover, although not widely reported in materials property optimization, AL has been shown to occasionally underperform RS in other domains [118, 119], which could arise due to inadequate surrogate model and/or acquisition function choices.
| Metric | Definition | Limitation |
| Discovery yield [33], [98] | Requires apriori knowledge of interesting materials in the pool | |
| Acceleration factor [94],[98] | Not (always) a fair baseline | |
| Enhancement factor [94],[98] | Not (always) a fair baseline | |
| Decision efficiency [33] | Dependence on pool quality | |
| Model performance [38], [104] | Error of surrogate model estimated on a (holdout) test set | Prone to overestimation due to active learning bias |
| Model uncertainty [86], [105] | Uncertainty of surrogate model | Prone to misleading due to unreliable uncertainties |
Table III. Evaluation metrics used for assessing AL performance for materials property optimization and discovery. is the active leaning iteration(s) and is the random sampling iterations. The figure of merit (FOM) is a general quantitative metric measuring the AL efficiency and could indicate the desired material property value, the number of promising materials in a target range, etc. The previous studies employed the metrics for performance evaluation of AL are indicated.
Another common evaluation criterion in materials property optimization and AL more broadly is tracking model performance over AL iterations. However, Borg et al. [98] observed that improvements in model accuracy do not necessarily correlate with better discovery rates of high-performing materials; in some cases, less accurate models identified promising candidates more effectively (Figure 3c). This suggests that the popular notion of “model is getting better” may not qualify as an accurate evaluation criteria for AL for materials discovery. Similarly, Koizumi et al.[120] demonstrated that the AL performance can vary substantially with changes in the material property, descriptor dimension and size of seed data even when using a fixed choice for surrogate model and sampling strategy. Kim et al. [121] emphasized that beyond the model and seed data, the quality of the candidate pool, particularly the fraction of promising materials, critically affects AL performance evaluation. They pointed out that materials property optimization depend on how likely candidates in the pool are to outperform those in the initial seed data and proposed metrics such as predicted fraction of improved candidates to quantify the pool quality. It is to be noted that all of these insights stem from post hoc studies using fully labeled datasets and are not based on AL workflows involving real-time experiments or simulations. Nevertheless, the use of diverse datasets spanning different properties, surrogate models, and acquisition functions, along with repeated trials to gather statistics suggests that these trends may hold more broadly, including for AL workflows involving real-time experiments and simulations.
In AL frameworks focused on multi-fidelity or multi-objective optimization, performance metrics must be adapted beyond those used in single-fidelity and single-objective tasks. In the multi-fidelity setting, data acquisition comes with differing costs and fidelities, making it essential to quantify not just accuracy, but also cost-efficiency. Jacobs et al.[122] demonstrated that low-fidelity data can improve optimization only when it is significantly cheaper (e.g. of high-fidelity cost) and partially available upfront. This underscores the need for performance metrics that account for both data cost and availability. While initial efforts exist in this direction [123, 124], general-purpose cost functions and robust sampling strategies for multi-fidelity materials optimization remain underdeveloped. In multi-objective AL, the aim is to identify candidate materials that perform well across multiple properties. Most studies have centered on Pareto optimization, where the goal is to approximate the Pareto front of non-dominated solutions [125, 126]. However, this approach is most effective when objectives are conflicting, which may not always be the case in materials science. For non-conflicting or correlated properties, scalarization (e.g. using weighted sums) may offer better performance, yet it introduces subjectivity in weight selection and risks missing diverse solutions. Furthermore, materials properties are often non-convex, unevenly distributed, and may vary in measurement cost, complicating both the optimization and its evaluation. These challenges highlight the pressing need for task-aware and cost-sensitive AL performance metrics that go beyond standard error or accuracy. Future efforts should aim to develop generalizable evaluation frameworks that reflect the realities of materials design—such as cost disparities, fidelity trade-offs, and property-specific constraints.
Algorithm 2 Proposed Hierarchical Sampling Scheme for General Active Learning Sampling
Require: Unlabeled candidate pool , representation , informativeness scoring function sample size , physical validity filter , cost function , representativeness threshold
Ensure: Selected samples satisfying multiple criteria
1: Step 1 — Physical & Feasibility Filtering: 2: ▷ Enforces physically meaningful, cost-feasible candidates 3: Step 2 — Informativeness Scoring: 4: Compute for each ▷ Model-based or model-free measure of expected information gain 5: Step 3 — Preselection of Top Candidates: 6: top- samples ranked by ▷ Limits to informative subset, avoids extreme outliers 7: Step 4 — Representativeness Filtering: 8: Compute representativeness score (e.g. local density or cluster-based weight) for 9: ▷ Ensures coverage 10: Step 5 — Diversity Selection (Sample Size ): 11: Select of size to maximize pairwise dissimilarity:
▷ can be distance-based or physical metric
12: return
Outlook
Active learning has shown substantial promise for accelerating materials simulation and discovery, yet its broader impact is still limited by unresolved challenges across the design and deployment of AL workflows. Overcoming these limitations requires development of better practices for the design and choice of various components of the AL pipeline. A key opportunity lies in the development of AL workflows that incorporate domain knowledge and practical constraints. For example, rather than focusing on a single sampling criterion, future strategies should increasingly adopt hierarchical schemes that jointly account for informativeness, representativeness, diversity, etc. an example for which is demonstrated in Algorithm 2. Such considerations are especially critical in materials science, where unphysical candidates or prohibitively expensive evaluations can quickly undermine the benefits of AL. Equally important is the need for clearly defining the set of hyperparametres (e.g. surrogate model, acquisition function choices) specific to AL and developing software tools that allow for their optimization. In addition, establishing standardized AL pipelines and conducting large-scale benchmarking across diverse materials datasets will be essential for deriving transferable guidelines and reducing reliance on ad hoc heuristics. Looking ahead, several emerging directions offer promising pathways to address persistent bottlenecks. The unreliability of surrogate models in early AL stages may be mitigated by leveraging pretrained or foundation models [127] trained on related materials properties, enabling more informed sampling with limited data. Reformulating AL as a sequential decision-making problem, such as through reinforcement learning, may further enable adaptive sampling policies in high-dimensional design spaces where conventional approaches struggle [128]. In parallel, recent advances in generative models open opportunities to couple AL with candidate generation, allowing efficient exploration of vast design spaces while iteratively steering generation toward high-performing regions [129, 130]. The progress in the design of AL workflows also needs a stronger integration between the materials and ML or communities as many advances relevant to AL such as uncertainty quantification developed by the later remain underexplored for AL for materials science applications. Bridging this gap through interdisciplinary efforts will be crucial for translating methodological advances into robust, scalable AL workflows capable of addressing the complexity and cost constraints of real-world materials applications.
Data Availability
No data is newly generated during this work.
Acknowledgement
A.S.N thanks Matthias Scheffler for helpful discussions, and Beate Paulus and the German Research Foundation (DFG) for support through the Walter-Benjamin Fellowship Program (project No. 540316537).
References
- [1]K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, Machine learning for molecular and materials science, Nature 559, 547 (2018).DOI
- [2]R. Vasudevan, G. Pilania, and P. V. Balachandran, Machine learning for materials design and discovery, Journal of Applied Physics 129 (2021).
- [3]S. Axelrod, D. Schwalbe-Koda, S. Mohapatra, J. Damewood, K. P. Greenman, and R. Gómez-Bombarelli, Learning matter: Materials design with machine learning and atomistic simulations, Accounts of Materials Research 3, 343 (2022).DOI
- [4]B. G. S. Grupt, R. K. Vasudevan, T. Potok, and S. V. Kalinin, A bridge for accelerating materials by design, NPJ Computational Materials 1, 1 (2015).
- [5]E. O. Pyzer-Knapp, J. W. J. T. Pera, P. W. Staar, S. Takeda, T. Laino, D. P. Sanders, J. Sexton, J. R. Smith, and A. Curioni, Accelerating materials discovery using artificial intelligence, high performance computing and robotics, npj Computational Materials 8, 84 (2022).
- [6]A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Scaling deep learning for materials discovery, Nature 624, 80 (2023).
- [7]W. F. Maier, K. Stoewe, and S. Sieg, Combinatorial and high-throughput materials science, Angewandte chemie international edition 46, 6016 (2007).
- [8]S. Curtarolo, G. L. Hart, M. B. Nardelli, N. Mingo, S. Sanvito, and O. Levy, The high-throughput highway to computational materials design, Nature materials 12, 191 (2013).DOI
- [9]P. Xu, X. Ji, M. Li, and W. Lu, Small data machine learning in materials science, npj Computational Materials 9, 42 (2023).
- [10]S. Bauer, P. Benner, T. Bereau, V. Blum, M. Boley, C. Carbogn, R. Catlow, G. Dehm, S. Eibl, R. Ernstofer, et al., Roadmap on data-centric materials science, Modelling and Simulation in Materials Science and Engineering (2024).
- [11]P. Karande, B. Gallagher, and T. Y.-J. Han, A strategic approach to machine learning for material science: how to tackle real-world challenges and avoid pitfalls, Chemistry of Materials 34, 7650 (2022).DOI
- [12]D. A. Cohn, Z. Ghahramani, and M. I. Jordan, Active learning with statistical models, Journal of artificial intelligence research 4, 129 (1996).
- [13]D. Cohn, L. Atlas, and R. Ladner, Improving generalization with active learning, Machine learning 15, 201 (1994).
- [14]K. Gubaev, E. V. Podryabinkin, and A. V. Shapeev, Machine learning of molecular properties: Locality and active learning, The Journal of chemical physics 148 (2018).
- [15]R. Jinnouchi, K. Miwa, F. Karsai, G. Kresse, and R. Asahi, On-the-fly active learning of interatomic potentials for large-scale atomistic simulations, The Journal of Physical Chemistry Letters 11, 6946 (2020).DOI
- [16]S. Nie, Y. Xiang, L. Wu, G. Lin, Q. Liu, S. Chu, and X. Wang, Active learning guided discovery of high entropy oxides featuring high h2-production, Journal of the American Chemical Society 146, 29325 (2024).
- [17]M. Suvarna, T. Zou, S. H. Chong, Y. Ge, A. J. Martín, and J. Pérez-Ramírez, Active learning streamlines development of high performance catalysts for higher alcohol synthesis, Nature Communications 15, 5844 (2024).DOI
- [18]K. Kang, T. A. Purcell, C. Carbognо, and M. Scheffler, Accelerating the training and improving the reliability of machine-learned interatomic potentials for strongly anharmonic materials through active learning, arXiv preprint arXiv:2409.11808 (2024).arxiv.org/abs/2409.11808
- [19]J. S. Smith, B. Nebgen, N. Lubbers, O. Isayev, and A. E. Roitberg, Less is more: Sampling chemical space with active learning, The Journal of chemical physics 148 (2018).
- [20]C. Verdi, F. Karsai, P. Liu, R. Jinnouchi, and G. Kresse, Thermal transport and phase transitions of zirconia by on-the-fly machine-learned interatomic potentials, npj Computational Materials 7, 156 (2021).arxiv.org/abs/2109.15002
- [21]T. Lookman, P. V. Balachandran, D. Xue, and R. Yuan, Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design, npj Computational Materials 5, 21 (2019).DOI
- [22]L. Bassman Oftelie, P. Rajak, R. K. Kalia, A. Nakano, F. Sha, J. Sun, D. J. Singh, M. Aykol, P. Huck, K. Persson, et al., Active learning for accelerated design of layered materials, npj Computational Materials 4, 74 (2018).
- [23]O. Ozbayram, D. Olsen, M. Annamaraju, A. E. Robertson, A. Venkatraman, S. R. Kalidindi, M. Zhou, and L. Graham-Brady, Batch active learning for microstructure–property relations in energetic materials, Mechanics of Materials 205, 105308 (2025).DOI
- [24]M. Hu, Q. Tan, R. Knibbe, M. Xu, G. Liang, J. Zhou, J. Xu, B. Jiang, X. Li, M. Rama Jayam, et al., Designing unique and high-performance al alloys via machine learning: Mitigating data bias through active learning, Computational Materials Science 244, 113204 (2024).DOI
- [25]H. Zhang, W. W. Chen, J. M. Rondinelli, and W. Chen, Et-al: entropy-targeted active learning for bias mitigation in materials data, Applied Physics Reviews 10 (2023).DOI
- [26]A. Jose, E. Devijver, N. Jakse, and R. Poloni, Informative training data for efficient property prediction in metal–organic frameworks by active learning, Journal of the American Chemical Society 146, 6134 (2024).
- [27]A. Wang, H. Liang, A. McDannald, I. Takeuchi, and A. G. Kusne, Benchmarking active learning strategies for materials optimization and discovery, Oxford Open Materials Science 2, itac006 (2022).arxiv.org/abs/2204.05838
- [28]Y. Tian, R. Yuan, D. Xue, Y. Zhou, X. Ding, J. Sun, and T. Lookman, Role of uncertainty estimation in accelerating materials development via active learning, Journal of Applied Physics 128 (2020).DOI
- [29]M. Kulichenko, B. Nebgen, N. Lubbers, J. S. Smith, K. Barros, A. E. Allen, A. Habib, E. Shinkle, N. Fedik, Y. W. Li, et al., Data generation for machine learning interatomic potentials and beyond, Chemical Reviews 124, 13681 (2024).DOI
- [30]K. Konyushkova, R. Sznitman, and P. Fua, Learning active learning from data, Advances in neural information processing systems 30 (2017).
- [31]C. Campbell, N. Cristianini, A. Smola, et al., Query learning with large margin classifiers, in ICML, Vol. 20 (2000) p. 0.
- [32]J.-N. Hwang, J. J. Choi, S. Oh, R. Marks, et al., Query-based learning applied to partially trained multilayer perceptrons, IEEE Transactions on Neural Networks 2, 131 (1991).
- [33]B. Rohr, H. S. Stein, D. Guevarra, Y. Wang, J. A. Haber, M. Aykol, S. K. Suram, and J. M. Gregoire, Benchmarking the acceleration of materials discovery by sequential learning, Chemical science 11, 2696 (2020).DOI
- [34]H. Khosravi, T. Olajire, A. S. Raihan, and I. Ahmed, A data driven sequential learning framework to accelerate and optimize multi-objective manufacturing decisions, Journal of Intelligent Manufacturing , 1 (2024).
- [35]D. Angluin, Queries and concept learning 2, 319 (1988).
- [36]L. Atlas, D. Cohn, and R. Ladner, Training connectionist networks with queries and selective sampling, Advances in neural information processing systems 2 (1989).
- [37]D. D. Lewis, A sequential algorithm for training text classifiers: Corrigendum and additional data, in Acm Sigir Forum, Vol. 29 (ACM New York, NY, USA, 1995) pp. 13–19.
- [38]K. Li, D. Persaud, K. Choudhary, B. DeCost, M. Greenwood, and J. Hattrick-Simpers, Exploiting redundancy in large materials datasets for efficient machine learning with less data, Nature Communications 14, 7283 (2023).DOI
- [39]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., Commentary: The materials project: A materials genome approach to accelerating materials innovation, APL materials 1 (2013).DOI
- [40]S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, M. Aykol, S. Rühl, and C. Wolverton, The open quantum materials database (oqmd): assessing the accuracy of dft formation energies, npj Computational Materials 1, 1 (2015).DOI
- [41]P. Lyngby and K. S. Thygesen, Data-driven discovery of 2d materials by deep generative models, npj Computational Materials 8, 232 (2022).DOI
- [42]C. J. Bartel, A. Trewartha, Q. Wang, A. Dunn, A. Jain, and G. Ceder, A critical examination of compound stability predictions from machine-learned formation energies, npj computational materials 6, 97 (2020).DOI
- [43]Y. Sheng, Y. Wu, J. Yang, W. Lu, P. Villars, and W. Zhang, Active learning for the power factor prediction in diamond-like thermoelectric materials, npj Computational Materials 6, 171 (2020).DOI
- [44]N. Bhatia, P. Rinke, and O. Krejčí, Leveraging active learning-enhanced machine-learned interatomic potential for efficient infrared spectra prediction, npj Computational Materials 11, 324 (2025).
- [45]L. Zhang, D.-Y. Lin, H. Wang, R. Car, and W. E, Active learning of uniformly accurate interatomic potentials for materials simulation, Physical Review Materials 3, 023804 (2019).DOI
- [46]B. Settles, Active learning literature survey, arXiv preprint arXiv:1010.1010 (2009).
- [47]Z. Wang and J. Ye, Querying discriminative and representative samples for batch mode active learning, ACM Transactions on Knowledge Discovery from Data (TKDD) 9, 1 (2015).
- [48]X. Li, D. Kuang, and C. X. Ling, Active learning for hierarchical text classification, in Pacific-Asia confer- ence on knowledge discovery and data mining (Springer, 2012) pp. 14–25.
- [49]B. Du, Z. Wang, L. Zhang, L. Zhang, W. Liu, J. Shen, and D. Tao, Exploring representativeness and informativeness for active learning, IEEE transactions on cybernetics 47, 14 (2015).
- [50]S.-J. Huang, R. Jin, and Z.-H. Zhou, Active learning by querying informative and representative examples, Advances in neural information processing systems 23 (2010).
- [51]V. Zaverkin, D. Holzmüller, I. Steinwart, and J. Kästner, Exploring chemical and conformational spaces by batch mode deep active learning, Digital Discovery 1, 605 (2022).DOI
- [52]K. Brinker, Incorporating diversity in active learning with support vector machines, in Proceedings of the 20th international conference on machine learning (ICML-03) (2003) pp. 59–66.
- [53]D. J. MacKay, Information-based objective functions for active data selection, Neural computation 4, 590 (1992).DOI
- [54]S. Dasgupta, Two faces of active learning, Theoretical computer science 412, 1767 (2011).
- [55]S. Farquhar, Y. Gal, and T. Rainforth, On statistical bias in active learning: How and when to fix it, arXiv preprint arXiv:2101.11665 (2021).arxiv.org/abs/2101.11665
- [56]C. Murray, J. U. Allingham, J. Antorán, and J. M. Hernández-Lobato, Addressing bias in active learning with depth uncertainty networks... or not, in I (Still) Can’t Believe It’s Not Better! Workshop at NeurIPS 2021 (PMLR, 2022) pp. 59–63.arxiv.org/abs/2112.06926
- [57]J. Hu, D. Liu, N. Fu, and R. Dong, Realistic material property prediction using domain adaptation based machine learning, Digital Discovery 3, 300 (2024).
- [58]L. E. Schultz, Y. Wang, R. Jacobs, and D. Morgan, A general approach for determining applicability domain of machine learning models, npj Computational Materials 11, 95 (2025).
- [59]X. Yang, Y. Liu, C. Mi, and X. Wang, Active learning kriging model combining with kernel-density-estimation-based importance sampling method for the estimation of low failure probability, Journal of Mechanical Design 140, 051402 (2018).
- [60]S. Xiong, J. Azimi, and X. Z. Fern, Active learning of constraints for semi-supervised clustering, IEEE Transactions on Knowledge and Data Engineering 26, 43 (2013).
- [61]D. Schwalbe-Koda, S. Hamel, B. Sadigh, F. Zhou, and V. Lordi, Model-free estimation of completeness, uncertainties, and outliers in atomistic machine learning using information theory, Nature Communications 16, 4014 (2025).
- [62]A. PA Subramanyam and D. Perez, Information-entropy-driven generation of material-agnostic datasets for machine-learning interatomic potentials, npj Computational Materials 11, 218 (2025).
- [63]A. G. Kusne, H. Yu, C. Wu, W. H. Zhang, J. Hattrick-Simpers, B. DeCost, S. Sarker, C. Oses, C. Toher, S. Curtarolo, et al., On-the-fly closed-loop materials discovery via bayesian active learning, Nature communications 11, 5966 (2020).
- [64]J. Moon, W. Beker, M. Siek, J. Kim, H. S. Lee, T. Hyeon, and B. A. Grzybowski, Active learning guides discovery of a champion four-metal perovskite oxide for oxygen evolution electrocatalysis, Nature Materials 23, 108 (2024).
- [65]J. Qi, T. W. Ko, B. C. Wood, T. A. Pham, and S. P. Ong, Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling, npj Computational Materials 10, 43 (2024).
- [66]V. Gkatsis, P. Maratos, C. Retsinas, G. Giannakopoulos, and P. Krokidas, Density-aware active learning for materials discovery: A case study on functionalized nanoporous materials, Physical Chemistry Chemical Physics (2025).
- [67]H. Jung, L. Sauerland, S. Stocker, K. Reuter, and J. T. Margraf, Machine-learning driven global optimization of surface adsorbate geometries, npj Computational Materials 9, 114 (2023).
- [68]H. Liu, B. Yucel, B. Ganapathysubramanian, S. R. Kalidindi, D. Wheeler, and O. Wodo, Active learning for regression of structure–property mapping: the importance of sampling and representation, Digital Discovery 3, 1997 (2024).
- [69]J. Sourati, M. Akcakaya, J. G. Dy, T. K. Leen, and D. Erdogmus, Classification active learning based on mutual information, Entropy 18, 51 (2016).
- [70]S. Wu, Z. Zhao, M. Yin, and H. Li, Fusing information entropy and similarity: A novel active learning strategy for chemical process fault classifications, Chemometrics and Intelligent Laboratory Systems 237, 104821 (2023).
- [71]Y. Kim and B. Shin, In defense of core-set: A density-aware core-set selection for active learning, in Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining (2022) pp. 804–812.
- [72]J. P. Janet, C. Duan, T. Yang, A. Nandy, and H. J. Kulik, A quantitative uncertainty metric controls error in neural network-driven chemical discovery, Chemical science 10, 7913 (2019).
- [73]Q. Li, N. Fu, S. S. Omee, and J. Hu, Md-hit: Machine learning for material property prediction with dataset redundancy control, npj Computational Materials 10, 245 (2024).
- [74]X. Zhan, H. Liu, Q. Li, and A. B. Chan, A comparative survey: Benchmarking for pool-based active learning., in IJCAI (2021) pp. 4679–4686.
- [75]A. P. Bartók, R. Kondor, and G. Csányi, On representing chemical environments, Physical Review B—Condensed Matter and Materials Physics 87, 184115 (2013).
- [76]S. S. Omee, N. Fu, R. Dong, M. Hu, and J. Hu, Structure-based out-of-distribution (ood) materials property prediction: a benchmark study, npj Computational Materials 10, 144 (2024).
- [77]J. Bi, Y. Xu, F. Conrad, H. Wiemer, and S. Ihlenfeldt, A comprehensive benchmark of active learning strategies with automl for small-sample regression in materials science, Scientific Reports 15, 37167 (2025).
- [78]A. Jose, J. P. A. de Mendonça, E. Devijver, N. Jakse, V. Monbet, and R. Poloni, Regression tree-based active learning, Data Mining and Knowledge Discovery 38, 420 (2024).
- [79]S. Kee, E. Del Castillo, and G. Runger, Query-by-committee improvement with diversity and density in batch active learning, Information Sciences 454, 401 (2018).
- [80]X. Yuan, M. Suvarna, J. Y. Lim, J. Pérez-Ramírez, X. Wang, and Y. S. Ok, Active learning-based guided synthesis of engineered biochar for co2 capture, Envi- ronmental Science & Technology 58, 6628 (2024).
- [81]K. Min and E. Cho, Accelerated discovery of novel inorganic materials with desired properties using active learning, The Journal of Physical Chemistry C 124, 14759 (2020).
- [82]F. Hase, L. M. Roch, C. Kreisbeck, and A. Aspuru-Guzik, Phoenix: a bayesian optimizer for chemistry, ACS central science 4, 1134 (2018).
- [83]M. Fronzi, O. Isayev, D. A. Winkler, J. G. Shapter, A. V. Ellis, P. C. Sherrill, N. A. Shepelin, A. Corletto, and M. J. Ford, Active learning in bayesian neural networks for bandgap predictions of novel van der waals heterostructures, Advanced Intelligent Systems 3, 2100080 (2021).
- [84]D. Xue, P. V. Balachandran, J. Hogden, J. Theiler, D. Xue, and T. Lookman, Accelerated search for materials with targeted properties by adaptive design, Nature communications 7, 1 (2016).
- [85]S. S. Hessmann, K. T. Schütt, N. W. Gebauer, M. Gastegger, T. Oguchi, and T. Yamashita, Accelerating crystal structure search through active learning with neural networks for rapid relaxations, npj Computational Materials 11, 44 (2025).
- [86]A. S. Nair, L. Foppa, and M. Scheffler, Materials-discovery workflow guided by symbolic regression for identifying acid-stable oxides for electrocatalysis, npj Computational Materials 11, 1 (2025).
- [87]B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. De Freitas, Taking the human out of the loop: A review of bayesian optimization, Proceedings of the IEEE 104, 148 (2015).
- [88]Y. Wu, A. Walsh, and A. M. Ganose, Race to the bottom: Bayesian optimisation for chemical problems, Digital Discovery 3, 1086 (2024).
- [89]A. Deshwal, C. M. Simon, and J. R. Doppa, Bayesian optimization of nanoporous materials, Molecular Systems Design & Engineering 6, 1066 (2021).
- [90]P. Honarmandi, V. Attari, and R. Arroyave, Accelerated materials design using batch bayesian optimization: A case study for solving the inverse problem from materials microstructure to process specification, Computational Materials Science 210, 111417 (2022).
- [91]J. K. Pedersen, C. M. Clausen, O. A. Kryasiak, B. Xiao, T. A. Batchelor, T. Löffler, V. A. Mints, L. Banko, M. Arenz, A. Savan, et al., Bayesian optimization of high-entropy alloy compositions for electrocatalytic oxygen reduction, Angewandte Chemie 133, 24346 (2021).
- [92]R. Yuan, Z. Liu, P. V. Balachandran, D. Xue, Y. Zhou, X. Ding, J. Sun, D. Xue, and T. Lookman, Accelerated discovery of large electrostrains in batio3-based piezo-electrics using active learning, Advanced materials 30, 1702884 (2018).
- [93]Y.-F. Lim, C. K. Ng, U. Vaitesswar, and K. Hippalgaonkar, Extrapolative bayesian optimization with gaussian process and neural network ensemble surrogate models, Advanced Intelligent Systems 3, 2100101 (2021).
- [94]Q. Liang, A. E. Gongora, Z. Ren, A. Tihonen, Z. Liu, S. Sun, J. R. Deneault, D. Bash, F. Mekki-Berrada, S. A. Khan, et al., Benchmarking the performance of bayesian optimization across multiple experimental materials science domains, npj Computational Materials 7, 188 (2021).
- [95]R. Moriconi, M. P. Deisenroth, and K. Sesh Kumar, High-dimensional bayesian optimization using low-dimensional feature spaces, Machine Learning 109, 1925 (2020).
- [96]C. Benjamins, E. Raponi, A. Jankovic, K. van der Blom, M. L. Santoni, M. Lindauer, and C. Doerr, Pi is back! switching acquisition functions in bayesian optimization, arXiv preprint arXiv:2211.01455 (2022).arxiv.org/abs/2211.01455
- [97]R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, Sisso: A compressed-sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates, Physical Review Materials 2, 083802 (2018).
- [98]C. K. Borg, E. S. Muckley, C. Nyby, J. E. Saal, L. Ward, A. Mehta, and B. Meredig, Quantifying the performance of machine learning models in materials discovery, Digital Discovery 2, 327 (2023).
- [99]J. Fromer, R. Wang, M. Manjrekar, A. Tripp, J. M. Hernández-Lobato, and C. W. Coley, Batched bayesian optimization by maximizing the probability of including the optimum, Journal of Chemical Information and Modeling (2025).
- [100]L. M. Roch, F. Häse, C. Kreisbeck, T. Tamayo-Mendoza, L. P. Yun, J. E. Hein, and A. Aspuru-Guzik, Chemos: orchestrating autonomous experimentation, Science Robotics 3, ea at5559 (2018).
- [101]F. Häse, M. Aldeghi, R. J. Hickman, L. M. Roch, M. Christensen, E. Liles, J. E. Hein, and A. Aspuru-Guzik, Olympus: a benchmarking framework for noisy optimization and experiment planning, Machine Learning: Science and Technology 2, 035021 (2021).
- [102]P. Pernot, Calibration in machine learning uncertainty quantification: beyond consistency to target adaptivity, APL Machine Learning 1 (2023).
- [103]Y. Hwang, W. Jo, J. Hong, and Y. Choi, Overcoming overconfidence for active learning, IEEE Access (2024).
- [104]L. Kavalsky, V. I. Hegde, E. Muckley, M. S. Johnson, B. Meredig, and V. Viswanathan, By how much can closed-loop frameworks accelerate computational materials discovery?, Digital Discovery 2, 1112 (2023).
- [105]L. Kavalsky, V. I. Hegde, B. Meredig, and V. Viswanathan, A multiobjective closed-loop approach towards autonomous discovery of electrocatalysts for nitrogen reduction, Digital discovery 3, 999 (2024).
- [106]D. Varivoda, R. Dong, S. S. Ome e, and J. Hu, Materials property prediction with uncertainty quantification: A benchmark study, Applied Physics Reviews 10 (2023).
- [107]K. Tran, W. Neiswanger, J. Yoon, Q. Zhang, E. Xing, and Z. W. Ulissi, Methods for comparing uncertainty quantifications for material property predictions, Machine Learning: Science and Technology 1, 025006 (2020).
- [108]C. J. Gruich, V. Madhavan, Y. Wang, and B. R. Goldsmith, Clarifying trust of materials property predictions using neural networks with distribution-specific uncertainty quantification, Machine Learning: Science and Technology 4, 025019 (2023).
- [109]G. Palmer, S. Du, A. Politowicz, J. P. Emory, X. Yang, A. Gautam, G. Gupta, Z. Li, R. Jacobs, and D. Morgan, Calibration after bootstrap for accurate uncertainty quantification in regression models, npj Computational Materials 8, 115 (2022).
- [110]Y. Chung, I. Char, H. Guo, J. Schneider, and W. Neiswanger, Uncertainty toolbox: an open-source library for assessing, visualizing, and improving uncertainty quantification, arXiv preprint arXiv:2109.10254 (2021).
- [111]C. Latanotis, S. Marelli, and B. Sudret, UqLab 2.0 and uqcloud: open-source vs. cloud-based uncertainty quantification, in SIAM Conference on Uncertainty Quantification (SIAM UQ 2022) (ETH Zurich, Institute of Structural Engineering, 2022).
- [112]A. Thomas-Mitchell, G. Hawe, and P. L. Popelier, Calibration of uncertainty in the active learning of machine learning force fields, Machine Learning: Science and Technology 4, 045034 (2023).
- [113]V. Korolev, I. Nevolin, and P. Protsenko, A universal similarity based approach for predictive uncertainty quantification in materials science, Scientific Reports 12, 14931 (2022).
- [114]J. Musieliwicz, J. Lan, M. Uyttendaele, and J. R. R. Kitchin, Improved uncertainty estimation of graph neural network potentials using engineered latent space distances, The Journal of Physical Chemistry C 128, 20799 (2024).
- [115]S. Zhong, D. R. Lambeth, T. K. Igou, and Y. Chen, Enlarging applicability domain of quantitative structure–activity relationship models through uncertainty-based active learning, ACS ES&T Engineering 2, 1211 (2022).
- [116]C. Sutton, M. Boley, L. M. Ghirnghelli, M. Rupp, J. Vreeken, and M. Scheffler, Identifying domains of applicability of machine learning models for materials science, Nature communications 11, 4428 (2020).
- [117]A. Palizhati, S. B. Torrisi, M. Aykol, S. K. Suram, J. S. Hummelshøj, and J. H. Montoya, Agents for sequential learning using multiple-fidelity data, Scientific reports 12, 4694 (2022).
- [118]A. I. Schein and H. Ungar, Active learning for logistic regression: an evaluation, Machine Learning 68, 235 (2007).
- [119]J. N. Fuhg, A. Fau, and U. Nackenhorst, State-of-the-art and comparative review of adaptive sampling methods for kriging, Archives of Computational Methods in Engineering 28, 2689 (2021).
- [120]A. Koizumi, G. Deffrennes, K. Terayama, and R. Tamura, Performance of uncertainty-based active learning for efficient approximation of black-box functions in materials science, Scientific Reports 14, 27019 (2024).
- [121]Y. Kim, E. Kim, E. Antono, B. Meredig, and J. Ling, Machine-learned metrics for predicting the likelihood of success in materials discovery, npj Computational Materials 6, 131 (2020).
- [122]R. Jacobs, P. E. Goins, and D. Morgan, Role of multi-fidelity data in sequential active learning materials discovery campaigns: case study of electronic bandgap, Machine Learning: Science and Technology 4, 045060 (2023).
- [123]V. Sabanzа-Gil, R. Barbano, D. P. Gutiérrez, J. S. Luterbacher, J. M. Hernández-Lobato, P. Schwaller, and L. Roch, Best practices for multi-fidelity bayesian optimization in materials and molecular research, arXiv preprint arXiv:2410.00544 (2024).
- [124]R. Magar and A. B. Farimani, Learning from mistakes: Sampling strategies to efficiently train machine learning models for material property prediction, Computational Materials Science 224, 112167 (2023).
- [125]V. Trinquet, M. L. Evans, C. J. Hargreaves, P.-P. De Breuck, and G.-M. Riganese, Optical materials discovery and design with federated databases and machine learning, Faraday Discussions 256, 459 (2025).
- [126]P. Xu, Y. Ma, W. Lu, M. Li, W. Zhao, and Z. Dai, Multi-objective optimization in machine learning assisted materials design and discovery, Journal of Materials Informatics 5, (2025).
- [127]I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kovács, J. Riebesell, X. R. Advincula, M. Asta, M. Avaylon, W. J. Baldwin, et al., A foundation model for atomistic materials chemistry, The Journal of chemical physics 163 (2025).
- [128]Y. Xian, X. Ding, X. Jiang, Y. Zhou, J. Sun, D. Xue, and T. Lookman, Unlocking the black box beyond bayesian global optimization for materials design using reinforcement learning, npj Computational Materials 11, 1 (2025).
- [129]R. Xin, E. M. Siriwardane, Y. Song, Y. Zhao, S.-Y. Louis, A. Nasiri, and J. Hu, Active-learning-based generative design for the discovery of wide-band-gap materials, The Journal of Physical Chemistry C 125, 16118 (2021).
- [130]M. Schwarting, L. Ward, N. Hudson, X. Yan, B. Blaiszik, S. Chaudhuri, E. Huerta, and I. Foster, Steering an active learning workflow towards novel materials discovery via queue prioritization, in 2025 IEEE International Conference on eScience (eScience) (IEEE, 2025) pp. 30–38.