Extracting reliable knowledge from unstructured materials literature remains a central bottleneck for data-driven and AI-enabled materials discovery. Large language models (LLMs) are reshaping this task by integrating multimodal document parsing, ontology-guided semantic grounding, structured extraction, and agentic verification into increasingly unified workflows. This review analyzes these developments through a Perception–Cognition–Action lens. At the perception layer, we examine how scientific document parsers, multimodal LLMs, table and chart readers, and optical chemical-structure-recognition systems convert visually rich papers into computable evidence. At the cognition layer, we discuss how ontologies and knowledge graphs constrain LLM outputs, support entity alignment, and reduce semantic ambiguity. At the action layer, we compare schema-based extraction, schema-free discovery, and agentic extraction as a control–coverage–autonomy spectrum rather than a simple succession of tools. We further argue that reliability is the decisive criterion for large-scale deployment, and synthesize failure modes, layered defenses, and evaluation protocols that connect source grounding, ontology constraints, physical verification, and human-in-the-loop review. By distinguishing demonstrated extraction capabilities from more speculative AI-scientist and self-driving-laboratory visions, this review provides a comparative and risk-aware account of how LLM-driven systems can produce evidence-linked, physically meaningful, and reusable materials knowledge.
Shuai Yang, Yimeng Wang, Qiong Tu et al.· Journal of Materials Informa...· 0 citations
The disparity between the complex structures of synthesized materials and their simplified computational models leads to deviations between theoretically calculated and experimental performance. To narrow this gap, we introduce the statistical descriptor φ, which is defined as the proportion of high-activity configurations in a given element combination. By considering the activity distribution of multiple structures rather than relying on a single model structure, φ can more accurately quantify macroscopic catalytic activity. Using the Seq-Equiformer model, a graph neural network we developed by augmenting EquiformerV2 with LSTM to capture dynamic structural changes during oxygen evolution reaction, we predict overpotentials for 250 million structures of 3d transition metal doped CoOOH. Based on these predictions, the value of φ for each element combination is calculated, and six optimal dopant combinations with the highest φ values are determined. For the leading MnFeNiCu combination, Bayesian optimization-driven AI experiments further optimize the elemental ratios. After only 40 experimental iterations, exploring 0.44% of the search space, the catalyst Mn0.07Fe0.09Ni0.14Cu0.01Co0.69OOH is identified, delivering an overpotential of 246.5 mV at 100 mA cm-2 and retaining 98.5% activity over 1000 h at 1 A cm-2. In validation, the statistical descriptor achieves 80% accuracy in identifying the top catalysts, a 30% improvement over single-structure screening, which evaluates the element combination based on the best configuration. The integration of statistical modeling, machine learning, and autonomous experimentation offers a powerful strategy to accelerate catalyst discovery and enhance prediction accuracy.
Chengbo Li, Mingzhe Li, N. Ran et al.· Journal of the American Chem...· 0 citations