Search

In the evolving landscape of computational and data-driven materials engineering, the integration of machine learning techniques has transformed traditional discovery paradigms into intelligent, autonomous systems. Materials informatics leverages vast datasets from high-throughput computations and multimodal sources to accelerate the design of novel materials with tailored properties. However, a conceptual gap persists in understanding the infrastructural roles of knowledge graphs and property predictors as competing yet complementary architectures for materials intelligence. Knowledge graphs offer relational representations that capture complex interdependencies among materials entities, enabling semantic querying and inference across disparate data modalities. In contrast, property predictors, often based on graph neural networks or deep learning models, focus on direct regression or classification of material attributes, prioritizing predictive accuracy over holistic system integration. This manuscript introduces a novel conceptual framework, termed the Dual-Infrastructure Materials Cognition (DIMC) model, which interprets the dynamic interplay between these infrastructures through layered computational workflows and feedback mechanisms. By examining representation learning, uncertainty quantification, and closed-loop discovery logics, the framework elucidates trade-offs in scalability, interpretability, and epistemic robustness. Implications for the field include enhanced steering of autonomous discovery systems, improved coupling of simulation and experimentation, and refined strategies for inverse materials design. Ultimately, this interpretive lens fosters a more cohesive ecosystem for materials intelligence, bridging isolated predictive tools with knowledge-centric infrastructures to advance data-driven innovation in materials science.

Knowledge graphs (KGs) have emerged as a pivotal infrastructure in computational and data-driven materials engineering, enabling structured representation, reasoning, and integration of heterogeneous data for accelerated discovery. By organizing materials data into interconnected entities and relationships, KGs facilitate advanced querying, inference, and machine learning applications across domains such as materials informatics, high-throughput computation, and inverse design. This review synthesizes recent advancements in KG construction from multimodal datasets, including text corpora, biomolecular integrations, and crystalline structures. We examine how graph neural networks and representation learning enhance molecular contrastive learning and pre-training frameworks for improved molecular representations. In the landscape of computational materials ecosystems, KGs support semantic integration and terminology standardization, bridging simulation and experiment through active learning systems and uncertainty quantification. Applications in autonomous laboratories highlight closed-loop discovery, where KGs enable dynamic knowledge propagation and event-sourced provenance management. We provide an original synthesis framing KGs as unifying backbones for data-model-experiment cycles, emphasizing systems-level integration over isolated tools. Challenges in scalability and interoperability are noted, with future directions toward hybrid human-AI workflows. This narrative underscores KGs' role in transforming materials discovery from empirical to predictive paradigms, fostering interdisciplinary convergence in materials science.