The field of materials science as a fundamental cornerstone for technological advancements and industrial transformation in society currently undergoes a profound paradigm shift propelled by artificial intelligence (AI). Conventional research and development (R&D) of materials, heavily reliant on empirical "trial-and-error" approaches and expert intuition, are inherently constrained by protracted R&D cycles and substantial financial costs. The advancement of the Materials Genome Initiative (MGI) has accumulated various large-scale scientific databases covering structure, property and process, but it remains a challenge that efficiently mines inherent law from these complex and multimodal massive data and integrates interdisciplinary information. The recent emergence and rapid evolution of large-scale foundation models offer a transformative solution to this critical bottleneck. Pre-trained on massive and diverse scientific corpora that include research literature, structural databases, and computational outputs, these models are evolving beyond mere pattern-recognition tools. They advance into sophisticated cognitive engines capable of decoding specialized material languages, integrate physical laws across multiple scales, and ultimately restructure the entire material discovery and optimization workflow. This review systematically examines a pivotal role of Materials Foundation Models (MFMs) in this AI-driven revolution. It delves into their core methodological foundations, showcases landmark applications across key sub-disciplines, and outlines the trajectory leading toward a future of fully intelligent and autonomous materials discovery.

The significance of MFMs is fundamentally rooted in their capacity to establish a unified and comprehensive evaluative framework, which is structured around their core underlying training paradigms. This framework is primarily built on three synergistic pillars, including large-scale pre-training on structural and textual corpora, supervised fine-tuning for specific property regressions and reinforcement learning from physical feedback, which aligns model outputs with inviolable thermodynamic and kinetic constraints. This review deconstructs their capabilities and limitations via analyzing models through this tripartite lens. A detailed architectural analysis of pioneering vertical domain models, such as MatMind for inorganic non-metallics and SteelBERT for metallurgy is given. We assess the specific mechanisms through that these models encode a deep domain-specific knowledge, including crystallographic symmetry rules, polymer chain topology dynamics, and solid-state reaction kinetics, into their latent representation spaces, thereby achieving a superior performance over general-purpose models. Furthermore, this review critically analyzes the sophisticated bridging mechanisms enabled by multi-agent frameworks. We examine how multi-agent systems like AtomAgents function as cognitive hubs, connecting and orchestrating high-throughput computational pipelines with experimental databases through unified natural language interfaces. These frameworks effectively manage the complex orchestration of diverse simulation tools and characterization equipment, thereby automating and accelerating the entire "design-simulate-characterize-optimize" cycle.

The research advances demonstrate a disruptive impact of these models across the materials R&D chain. MFMs significantly enhance design efficiency and prediction accuracy via suppressing "scientific hallucinations" through the integration of explicit physical rules and attention visualization techniques. In the realm of crystal discovery, the GNoME model effectively expands the library of stable inorganic crystals by ten times through an active learning closed-loop. For structure generation, models like CrystaLLM demonstrate an ability to capture atomic-scale symmetry breaking during phase stability predictions via deconstructing attention distributions over crystal descriptors. In specialized sectors, vertical models such as PolyLLM and ChipNeMo effectively bridge knowledge gaps via embedding domain-specific heuristics, while in the metallurgical sector, SteelBERT and PDGPT achieve a high accuracy in predicting steel properties via correlating process parameters with microstructural evolution. Moreover, multimodal frameworks such as MatterChat bridges a gap between natural language and atomic-scale features, where attention weights are automatically assigned to active-site orbitals, which is highly consistent with coordination field theory. The integration of agentic workflows further demonstrates that these foundation models can manage the entire cycle with an unprecedented precision.