Data-driven active learning approaches for accelerating materials discovery

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: Materials discovery is a cornerstone of modern technological advancement, yet it remains constrained by traditional trial-and-error paradigms and the inherent bias of human intuition. Artificial intelligence (AI) has emerged as a transformative tool in materials science by effectively modeling structure-property relationships. Despite substantial efforts to enhance model expressiveness, data efficiency remains an equally critical challenge, given the limited availability of experimental and computational resources. Active learning (AL), as a data-driven machine learning paradigm, has shown great promise for discovering novel materials and enabling the efficient navigation of vast materials spaces. In this review, we follow the evolution of sampling strategy design techniques in AL, from Bayesian optimization to advanced deep learning-based strategies. We then highlight how AL enhances data efficiency across various data regimes, ranging from task-specific settings with limited data to the development of general-purpose datasets and large-scale models. We further provide a systematic overview of AL applications throughout the materials research pipeline, including computational simulation, composition and structural design, process optimization, and self-driving laboratory systems. Finally, we pinpoint key challenges and future perspectives of AL in materials discovery.

Introduction: The Data Efficiency Challenge

The search for new materials—such as superconductors, catalysts, or high-performance alloys—is a "needle in a haystack" problem. With a chemical search space estimated at $10^{60}$ possible molecules, traditional discovery relies heavily on human intuition and costly, time-consuming experiments. While standard AI has helped model structure-property relationships, it is notoriously "data-hungry." In materials science, data is a luxury: a single high-fidelity simulation (DFT) or lab synthesis can take days. Active Learning (AL) addresses this data efficiency gap by strategically selecting only the most "informative" data points for expensive evaluation, building accurate models with the smallest possible number of data points.

Core Components of Active Learning

The evolution of AL in materials science spans from traditional statistical methods to modern deep learning architectures. The framework generally revolves around three pillars: the model, the data, and the acquisition function.

The iterative Active Learning loop and common sampling scenarios: stream-based, pool-based, and membership query synthesis.
  • Traditional AL: Primarily centered on Bayesian Optimization (BO), using probabilistic surrogate models (like Gaussian Processes) to predict properties and acquisition functions to balance exploitation of known high-performing regions with exploration of uncertain ones.
  • Deep Active Learning (DAL): Integrates Deep Neural Networks to handle high-dimensional data. Techniques like Monte Carlo Dropout and Bayesian Neural Networks allow the AI to estimate its own uncertainty.
  • Sampling Strategies: Models choose the next data point based on uncertainty (picking points where the model is most confused), distribution (ensuring diverse coverage of the structural landscape), or hybrid approaches.

The Active Learning Workflow

The AL process is a closed-loop system designed to iteratively improve model performance. It begins with a small set of labeled data, trains a surrogate model, and uses an active learning strategy to query an "Oracle" (either an experimentalist or a computational simulation) for new labels.

The iterative Active Learning loop and common sampling scenarios: stream-based, pool-based, and membership query synthesis.

There are three primary sampling scenarios:

  1. Stream-based Selective Sampling: The model decides in real-time whether to query or discard a data point from a continuous stream.
  2. Pool-based Sampling: The model selects the best query from a large, pre-defined pool of unlabeled candidates (e.g., from literature or computational repositories).
  3. Membership Query Synthesis: The model generates a completely new query de novo, creating candidates that may not exist in any current database.

Key Insights and Results

Active Learning is no longer just a tool for specific tasks; it is becoming the engine for building "foundation models." By using AL to select diverse, non-equilibrium structures, researchers can train universal interatomic potentials that generalize across the entire periodic table.

  • Massive Acceleration: In simulations of ionic conductivity, AL-driven frameworks achieved a 19,000-fold acceleration relative to traditional density functional theory (DFT).
  • Efficiency Gains: In catalyst discovery, AL identified high-performance compositions by evaluating less than 5% of the initial candidate pool.
  • Scale of Discovery: Projects like DeepMind’s GNoME used an AL "flywheel" to identify 2.2 million stable crystals, expanding the number of known stable materials by a factor of ten.
  • LLM Frontier: Large Language Models (LLMs) are now being used as reasoning agents, reaching optimal candidates using less than 30% of the data required by traditional models.

Limitations and Future Perspectives

Despite its success, several challenges remain. The "Cold-Start Paradox"—where AL can perform worse than random sampling in the initial stages—requires better strategies for picking seed data. Furthermore, the field lacks standardized benchmarks to compare different AL configurations across diverse material classes like polymers versus high-entropy alloys.

Future directions include Physics-Informed AL, which integrates fundamental laws (thermodynamics, symmetry) directly into the sampling process, and the continued development of Self-Driving Laboratories (SDLs). These closed-loop labs represent the ultimate realization of AL, where AI plans experiments, robots execute synthesis, and results are fed back in real-time without human intervention.

Summary updated

Abstract

Materials discovery is a cornerstone of modern technological advancement, yet it remains constrained by traditional trial-and-error paradigms and the inherent bias of human intuition. Artificial intelligence (AI) has emerged as a transformative tool in materials science by effectively modeling structure-property relationships. Despite substantial efforts to enhance model expressiveness, data efficiency remains an equally critical challenge, given the limited availability of experimental and computational resources. Active learning (AL), as a data-driven machine learning paradigm, has shown great promise for discovering novel materials and enabling the efficient navigation of vast materials spaces. In this review, we follow the evolution of sampling strategy design techniques in AL, from Bayesian optimization to advanced deep learning-based strategies. We then highlight how AL enhances data efficiency across various data regimes, ranging from task-specific settings with limited data to the development of general-purpose datasets and large-scale models. We further provide a systematic overview of AL applications throughout the materials research pipeline, including computational simulation, composition and structural design, process optimization, and self-driving laboratory systems. Finally, we pinpoint key challenges and future perspectives of AL in materials discovery.
BibTeX
@misc{chen2026data,
  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 = {{Data-driven active learning approaches for accelerating materials discovery}},
  year = {2026},
  eprint = {2601.06971},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.06971},
}
Figure