A survey of active learning in materials science: Data-driven paradigm for accelerating the research pipeline

Jiaxin Chen, Tianjiao Wan, Hui Geng, Liang Xiong, Guohong Wang, Yihan Zhao, Linfeng Deng, Susu Fang, Zheng Luo, Huaimin Wang, Shanshan Wang, Kele Xu

Abstract: The exploration of materials composition, structure, and processing spaces is constrained by high dimensionality and the cost of data acquisition. While machine learning has supported property prediction and design, its effectiveness depends on labeled data, which remains expensive to generate via experiments or high-fidelity simulations. Improving data efficiency is thus a central concern in materials informatics. Active learning (AL) addresses this by coupling model training with adaptive data acquisition. Instead of static datasets, AL iteratively prioritizes candidates based on uncertainty, diversity, or task-specific objectives. By guiding data collection under limited budgets, AL offers a structured approach to decision-making, complementing physical insight with quantitative measures of informativeness. Recently, AL has been applied to computational simulation, structure optimization, and autonomous experimentation. However, the diversity of AL formulations has led to fragmented methodologies and inconsistent assessments. This Review provides a concise overview of AL methods in materials science, focusing on their role in improving data efficiency under realistic constraints. We summarize key methodological principles, representative applications, and persistent challenges, aiming to clarify the scope and limitations of AL as a practical tool within contemporary materials informatics.

Introduction: The Data Efficiency Crisis

In the pursuit of advanced energy storage, high-performance catalysts, and structural alloys, the primary bottleneck is the high cost of experimental and computational validation. Traditional materials discovery relies on a slow, iterative "trial-and-error" process. While machine learning (ML) has accelerated property prediction, these models are often "data-starved." High-fidelity tools like Density Functional Theory (DFT) are computationally intensive, and physical laboratory synthesis is even more resource-demanding.

Active Learning (AL) has emerged as a critical solution to this data efficiency crisis. By transforming the discovery process into an intelligent, iterative loop, researchers can navigate vast chemical spaces—the effectively infinite combinations of elements and structures—while performing only a fraction of the traditional experiments. AL acts as a "data flywheel," strategically selecting the most informative data points to label next, thereby balancing the exploitation of known promising regions with the exploration of high-uncertainty domains.

Methodological Framework

The core of the AL paradigm is the closed-loop workflow, which iteratively refines a surrogate model's understanding of the material space. This process typically follows four distinct stages:

  1. Surrogate Modeling: A model is trained on an initial, often small, labeled dataset.
  2. Candidate Selection: An acquisition function evaluates an unlabeled pool of candidates to identify the most "valuable" points based on criteria like uncertainty or expected improvement.
  3. Oracle Labeling: The selected candidates are evaluated by an "Oracle"—either a high-fidelity simulation (e.g., DFT) or a physical experiment.
  4. Model Update: The new data is integrated into the training set, and the loop repeats.
The Active Learning closed-loop workflow and sampling strategies, including stream-based, pool-based, and membership query synthesis.

Evolution of AL Architectures

The field has transitioned through two primary methodological eras:

  • Traditional Active Learning: Primarily utilizes Bayesian Optimization (BO) with Gaussian Processes (GPs). These models are favored for their inherent ability to provide mathematical uncertainty scores, which are essential for acquisition functions like Expected Improvement (EI) or Upper Confidence Bound (UCB).
  • Deep Active Learning (DAL): To handle the increasing scale of materials data, researchers are integrating Deep Learning (Neural Networks) with AL. Since standard neural networks can be overconfident, techniques such as Ensembles, Monte Carlo Dropout, and Evidential Deep Learning are employed to extract reliable uncertainty signals.

Taxonomy of Active Learning in Materials

The implementation of AL requires careful selection of models, data strategies, and acquisition objectives. As shown in the figure below, the framework encompasses a variety of approaches tailored to specific research goals, ranging from simple screening of known spaces to the extrapolation into entirely new chemical territories.

Taxonomy of active learning components, highlighting the relationships between models, data acquisition strategies, and acquisition functions.
  • The Cold-Start Paradox: AL can initially underperform compared to random sampling. Strategies like space-filling (Sobol sequences) or domain-knowledge-guided initialization are critical to anchoring the model before the iterative loop begins.
  • Extrapolation vs. Screening: While AL is highly effective for optimizing known materials (screening), its most transformative potential lies in extrapolation—pushing the boundaries of known chemistry to discover novel material classes that elude human intuition.
  • Large Language Models (LLMs): An emerging trend involves using LLMs as surrogate models. Their ability to "reason" through experimental parameters in natural language has shown promise, particularly in the early stages of discovery where structured data is sparse.

Impact and Results

The application of AL has yielded significant breakthroughs across materials informatics:

  • Efficiency Gains: In catalyst discovery, AL-guided loops have identified optimal compositions using less than 5% of the total candidate pool.
  • Computational Acceleration: The SurFF framework achieved a 100,000-fold acceleration in predicting surface energy and morphology for intermetallic crystals.
  • Discovery at Scale: DeepMind’s GNoME project utilized AL to identify 2.2 million stable crystals, expanding known materials databases by an order of magnitude—a feat impossible via brute-force methods.
  • Small-Data Success: Researchers successfully identified two new "Invar" high-entropy alloys using only 17 physical experiments, demonstrating AL's utility even in resource-constrained laboratory settings.

Challenges and Future Directions

Despite its success, several hurdles remain for the widespread adoption of AL:

  1. Benchmarking: The field lacks standardized benchmarks (an "ImageNet" for materials), making it difficult to compare the efficacy of different AL strategies across diverse datasets.
  2. Physics-Informed Architectures: Most AL models remain "physics-blind." Future research must focus on embedding physical constraints, such as symmetry and conservation laws, directly into model architectures.
  3. Human-in-the-Loop Integration: While the goal is fully autonomous "Self-Driving Labs," robust tools are needed to allow human experts to intervene and guide the AI when it encounters non-physical or irrelevant regions of the search space.
  4. Accessibility: Democratizing AL requires user-friendly, materials-specific software packages (e.g., Bgolearn) that allow bench scientists to implement these complex workflows without extensive coding expertise.

Summary updated

Abstract

The exploration of materials composition, structure, and processing spaces is constrained by high dimensionality and the cost of data acquisition. While machine learning has supported property prediction and design, its effectiveness depends on labeled data, which remains expensive to generate via experiments or high-fidelity simulations. Improving data efficiency is thus a central concern in materials informatics. Active learning (AL) addresses this by coupling model training with adaptive data acquisition. Instead of static datasets, AL iteratively prioritizes candidates based on uncertainty, diversity, or task-specific objectives. By guiding data collection under limited budgets, AL offers a structured approach to decision-making, complementing physical insight with quantitative measures of informativeness. Recently, AL has been applied to computational simulation, structure optimization, and autonomous experimentation. However, the diversity of AL formulations has led to fragmented methodologies and inconsistent assessments. This Review provides a concise overview of AL methods in materials science, focusing on their role in improving data efficiency under realistic constraints. We summarize key methodological principles, representative applications, and persistent challenges, aiming to clarify the scope and limitations of AL as a practical tool within contemporary materials informatics.
BibTeX
@misc{chen2026survey,
  author = {Jiaxin Chen and Tianjiao Wan and Hui Geng and Liang Xiong and Guohong Wang and Yihan Zhao and Linfeng Deng and Susu Fang and Zheng Luo and Huaimin Wang and Shanshan Wang and Kele Xu and Kele Xu},
  title = {{A survey of active learning in materials science: Data-driven paradigm for accelerating the research pipeline}},
  year = {2026},
  eprint = {2601.06971},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.06971},
}
Figure