Search

Generative models have emerged as transformative tools in materials science, enabling the inverse design of novel materials with tailored properties by learning from vast datasets of structures and compositions. This review synthesizes recent advancements in generative approaches, including variational autoencoders, generative adversarial networks, diffusion models, and large language models. It highlights their conceptual capabilities for accelerating discovery while addressing scientific limits such as data scarcity, synthesizability, and interpretability. By examining applications in inorganic crystals, organic molecules, and energy materials, we delineate how these models bridge computational efficiency with experimental validation, yet face challenges in generalizability and physical fidelity. Future directions emphasize hybrid physics-informed architectures and closed-loop automation to overcome current barriers and unlock sustainable materials innovation.

In materials science, the relationship between microstructure and material properties underpins rational design and performance optimization. Still, due to the complexity, heterogeneity, and multiscale nature of microstructural data, it is difficult to recognize. Electron microscopy provides rich visual access to microstructures. Still, existing analysis approaches rely heavily on manual interpretation or supervised machine learning, both of which are limited by the scarcity of annotations and limited generalizability. This paper presents the hierarchical invariant microstructure representation (HIMR) framework as a purely theoretical contribution to self-supervised representation learning for analyzing the microstructure of electron microscopy images. Rather than proposing an algorithm or empirical pipeline, HIMR provides a conceptual framework for learning, structuring, and relating microstructural information to material properties without labeled data. This framework conceptualizes microstructures as hierarchically organized latent representations, where physically meaningful features emerge through invariance-driven self-supervision and scale-aware aggregation. By integrating principles from representation learning, self-supervised paradigms, and materials physics, HIMR addresses foundational challenges, including imaging variability, scale entanglement, and the disconnect between the learned properties and physically interpretable property reasoning. Central to the framework is the alignment of the learned representation manifolds with property spaces governed by physical laws, enabling interpretable and theoretically grounded microstructure–property reasoning. By articulating explicit theoretical commitments regarding hierarchy, invariance, interpretability, and epistemic restraint, this work advances a framework-level understanding of self-supervised learning in materials science. As a result, HIMR provides a durable conceptual foundation for autonomous, data-efficient, and physically grounded analysis in AI-driven materials discovery and engineering.

The integration of artificial intelligence (AI) and machine learning (ML) into materials science has fundamentally transformed how material properties are predicted, analyzed, and understood. While early data-driven approaches emphasized predictive accuracy and high-throughput screening, recent advances are increasingly focusing on interpretability and explainability, enabling AI models to contribute to mechanistic scientific insight rather than functioning as opaque black boxes. This study examines the evolution of interpretable AI in materials science and highlights the transition from property prediction to explanation-driven understanding of structure–property relationships. In this thesis, we investigate the progress in machine learning frameworks that operate with limited or implicit structural information, alongside the growing use of explainable AI (XAI) techniques to uncover physically meaningful descriptors, atomic-scale interactions, and microstructural drivers of material behavior. Methods such as graph-based learning, attention mechanisms, feature attribution, and uncertainty-aware modeling are discussed for their ability to improve model reliability, expose data bias, and guide hypothesis generation. Representative applications across alloys, perovskites, organic semiconductors, and ferroelectric materials demonstrate how interpretable models have revealed governing mechanisms spanning atomic, mesoscopic, and macroscopic length scales. Beyond individual case studies, this study examines persistent challenges in interpretable materials AI, including data quality, generalizability, explanation stability, and computational overhead. We argue that interpretability is not merely an auxiliary feature but a prerequisite for trustworthy and scientifically helpful AI in materials research. By synthesizing recent methodological and application-driven advances, this review positions interpretable AI as a critical enabler of mechanism-oriented discovery, experimental validation, and theory development, ultimately advancing AI from a predictive accelerator to an integral partner in scientific understanding.

The integration of artificial intelligence (AI) into materials science represents a profound epistemic shift, challenging longstanding assumptions about the nature and validation of scientific knowledge. Rather than merely accelerating computational tasks, AI reconfigures the epistemological landscape by generating outputs that blur the boundaries between data, inference, and insight. This paper diagnoses a central problem: “knowledge inflation,” where AI’s predictive prowess is prematurely equated with genuine understanding, leading to overconfidence in materials-related decisions. Tensions arise between the opacity of AI-driven predictions and the demands for explanation and mechanistic clarity inherent to materials science, where structure-property relationships and causal processes have traditionally grounded epistemic warrant. Such discrepancies risk undermining the reliability of knowledge claims in domains like alloy design and sustainable material selection. To address this, we propose a reframed epistemic framework tailored to materials AI, centered on actionability as the criterion for knowledge: what can be responsibly acted upon in practical contexts. This includes a novel typology distinguishing epistemic categories of AI outputs, from mere predictive signals to robust decision warrants, with conditions for elevation between them. By emphasizing responsibility and scope, this reframing aims to safeguard epistemic integrity while harnessing AI’s potential.

The rise of artificial intelligence (AI) in materials science has highlighted a profound epistemic tension. While AI models excel in predictive accuracy, they often fail to provide mechanistic insights into materials behavior, raising questions about whether such predictions constitute genuine scientific understanding. This tension is particularly acute in materials science, where complex phenomena like phase transitions, defect dynamics, and property emergence demand not only forecasting but also explanatory depth to inform reliable design and innovation. Equating prediction with understanding risks epistemic overreach, potentially leading to unwarranted confidence in AI outputs and hindering progress in fields requiring causal knowledge, such as sustainable materials development. This paper proposes a novel theoretical framework that redefines “understanding” in AI-driven materials research as a multi-layered epistemic construct, distinguishing predictive success from mechanistic insight and actionable knowledge. The framework introduces epistemic validity conditions, interpretive constraints, and decision contexts for evaluating AI contributions, emphasizing alignment with physical principles and the avoidance of semantic inflation. By synthesizing recent literature, it addresses conceptual gaps in current approaches and advocates responsible inference that integrates predictive power with explanatory rigor. This contribution advances philosophical foundations for AI in materials science, fostering more robust, trustworthy scientific practices without empirical validation claims.

The rapid integration of artificial intelligence (AI) into materials science has enabled unprecedented predictive capabilities across a wide range of properties and structures. However, the predominantly black-box nature of these models limits their epistemic role, confining them largely to correlative tools rather than instruments capable of supporting genuine scientific reasoning. This conceptual manuscript introduces a novel theoretical framework that delineates a structured pathway for validating AI systems as materials reasoning tools. Drawing on recent advances in explainable and interpretable AI, as well as philosophical accounts of scientific reasoning, the framework articulates a progressive sequence of validation stages: establishing transparency and interpretability, extracting mechanistically meaningful explanations, assessing reasoning fidelity through inferential behavior, and integrating AI systems as instruments within the broader scientific knowledge cycle. The approach is deliberately architecture-agnostic and avoids empirical prescriptions, focusing instead on the conceptual and epistemic conditions required for scientific legitimacy. By explicitly bridging predictive performance with explanatory depth, inferential robustness, and alignment with physical theory, the proposed pathway reframes how success in materials AI is evaluated. It provides a foundation for distinguishing advanced predictive engines from systems capable of contributing to hypothesis generation, theory refinement, and cumulative understanding. In doing so, the framework addresses persistent barriers to the acceptance of AI as a scientific partner in materials research. It offers a principled basis for future methodological and evaluative developments.

The rapid integration of artificial intelligence (AI) into materials science marks a profound shift in how materials are discovered, characterized, and optimized. Rather than functioning merely as a computational aid, AI increasingly operates as an epistemic instrument that reshapes scientific workflows, decision-making practices, and notions of explanation within the field. This narrative review examines the conceptual foundations underpinning applied AI in materials science, with a particular focus on core definitions, implicit and explicit assumptions, and unresolved debates that continue to shape the domain. Key AI paradigms—including supervised, unsupervised, and reinforcement learning—are situated within materials-specific contexts such as property prediction, structure–property mapping, and autonomous experimentation. The review critically interrogates foundational assumptions regarding data quality, representativeness, generalization, and model transferability, highlighting how these assumptions condition both the successes and failures of AI-driven materials research. Persistent debates surrounding interpretability, epistemic trust, ethical responsibility, and environmental sustainability are synthesized from recent literature published. By articulating both the transformative potential and the conceptual limitations of applied AI, this review underscores the necessity of rigorous validation, transparent reasoning, and interdisciplinary collaboration to ensure that AI contributes robustly and responsibly to materials innovation.

The integration of artificial intelligence (AI) and machine learning (ML) into materials science has revolutionized the discovery, design, and optimization of new materials, enabling accelerated predictions of properties and behaviors previously unattainable with traditional methods. However, the “black-box” nature of many advanced AI models poses significant challenges, including a lack of transparency that hinders scientific understanding, trust, and practical adoption in materials research. This narrative review explores the concept of interpretability in materials AI, focusing on what constitutes an “explanation” and how it should be conceptually evaluated. Drawing from recent advancements in explainable AI (XAI), we delineate definitions of explanations tailored to materials informatics, emphasizing their role in bridging computational predictions with physical insights. We examine thematic aspects such as intrinsic versus post-hoc interpretability methods, the multidimensional nature of explanations (e.g., local vs. global, feature-based vs. mechanistic), and conceptual frameworks for evaluation, including criteria like fidelity, comprehensibility, robustness, and domain-specific relevance. By synthesizing the literature, we highlight how explanations can enhance materials discovery across alloy design, catalyst development, and polymer engineering, while addressing gaps in current evaluation practices. The review underscores the need for standardized conceptual metrics that go beyond quantitative benchmarks to incorporate qualitative, human-centered assessments in materials science contexts. Ultimately, this work aims to guide researchers toward developing interpretable AI systems that not only predict but also elucidate underlying material phenomena, fostering a more insightful and ethical application of AI in materials innovation.

Physics-guided machine learning (PGML) has emerged as a hybrid paradigm in materials science, integrating domain knowledge with data-driven methods to enhance predictive accuracy and generalizability. Conventional approaches typically embed physical principles as soft inputs—either through loss-function regularization or auxiliary features—allowing violations during optimization. This manuscript advances a conceptual reframing in which physics operates as a hard constraint on the model’s hypothesis space rather than as an additive input. By restricting permissible functional forms, symmetries, and conservation relations a priori, the framework enforces physical consistency at the architectural level, altering the interaction dynamics between data and prior knowledge. The reframing yields systems-level insights into epistemic trade-offs: reduced reliance on large datasets, improved extrapolation beyond training regimes, and inherent satisfaction of thermodynamic or mechanical invariants critical to materials behavior. Analytical implications include feedback structures that couple data refinement to constraint satisfaction, revealing emergent robustness in multiscale modeling. This perspective addresses persistent challenges in materials science, such as sparse experimental data and complex microstructure-property relationships, without resorting to empirical validation. The contribution lies in reinterpreting PGML’s epistemic foundation, steering future developments toward constraint-centric designs that prioritize physical fidelity over post-hoc penalization.

The integration of artificial intelligence (AI) into materials science has transformed the landscape of material discovery and design, enabling accelerated exploration of vast compositional spaces and property predictions. However, this advancement introduces complex error dynamics that permeate the entire AI pipeline, from data curation to model deployment. This conceptual manuscript develops a novel system-level model to interpret error propagation within materials AI workflows, emphasizing interaction dynamics and feedback structures rather than empirical validation. Drawing on recent literature, the synthesis reveals fragmented understandings of error sources, such as data inconsistencies and algorithmic biases, and their cascading effects across pipeline stages. The proposed framework conceptualizes the pipeline as an interconnected system where errors manifest through amplification, mitigation, and transformation mechanisms, informed by epistemic considerations and trade-off analyses. Analytical implications highlight how these dynamics influence interpretability and reliability in materials innovation, while ethical reasoning underscores the need for holistic oversight. By integrating conceptual interpretations from uncertainty quantification and systems theory, this model offers insights into steering logics that balance precision with robustness, fostering a deeper understanding of AI’s role in advancing materials science without relying on testable claims or experimental data.

The integration of artificial intelligence (AI) into materials science represents a paradigm shift in how scientific creativity is manifested and harnessed. This conceptual paper develops a novel theoretical framework for understanding AI-mediated hypothesis generation, emphasizing its role in enhancing scientific creativity within materials discovery and design. Traditional hypothesis generation in materials science relies on human intuition, empirical observation, and theoretical deduction, often constrained by cognitive limitations and the vast complexity of material systems. AI, through machine learning algorithms and generative models, augments this process by enabling rapid pattern recognition, simulation of hypothetical scenarios, and exploration of uncharted chemical spaces. The proposed framework, termed the symbiotic creativity cycle (SCC), posits a dynamic interplay between human and AI agents, where AI serves as a cognitive amplifier, facilitating divergent exploration and convergent refinement of hypotheses. This cycle incorporates iterative feedback loops that integrate domain knowledge with data-driven insights, fostering emergent creativity that transcends individual capabilities. Key elements includeAI’s ability to handle multidimensional data, predict material properties, and generate novel conceptual blends. The framework highlights potential applications for accelerating discoveries in advanced alloys, nanomaterials, and energy storage materials, while addressing challenges such as interpretability and ethical integration. By reconceptualizing scientific creativity as a hybrid human-AI endeavor, this paper lays the foundation for future theoretical developments and practical applications in applied artificial intelligence for materials science. Ultimately, AI-mediated hypothesis generation promises to democratize innovation, enabling more efficient navigation of the materials design landscape.

Transfer learning has emerged as a pivotal strategy in materials science, enabling the reuse of knowledge from data-rich domains to inform predictions in data-scarce contexts, thereby accelerating discovery across alloy design, nanomaterials, and functional compounds. Despite its growing adoption, the effectiveness of transfer learning remains contingent on subtle boundary conditions that delineate productive knowledge integration from ineffective or counterproductive transfer. This conceptual paper develops a theoretical framework to interpret these boundaries by examining interaction dynamics between source and target domains in materials contexts. It explores how mismatches in representational hierarchies—such as between atomic-scale and macroscopic descriptions—disrupt knowledge flow and yield distorted predictive outcomes. Systems-level analysis reveals trade-offs in model adaptability, where reliance on pre-trained representations may obscure emergent properties specific to target materials. Ethical considerations further highlight the risks of bias propagation from simulated to experimental domains, with implications for research prioritization and resource allocation. By integrating perspectives from materials informatics and complexity theory, the framework articulates steering logics to mitigate transfer failures through adaptive feature alignment. This work advances conceptual understanding of transfer learning limitations and provides interpretive guidance for future AI integration in materials science, without empirical validation.

The integration of physical principles into machine learning (ML) frameworks has emerged as a transformative approach in materials science, addressing the limitations of purely data-driven models by incorporating domain knowledge to enhance predictive accuracy, generalizability, and interpretability. This narrative review explores the conceptual taxonomies of physics-integrated ML methods, their applications in materials discovery and design, and the associated challenges in data bias and ethical considerations. Drawing on recent peer-reviewed literature, we classify physics-integration strategies such as physics-informed neural networks (PINNs), hybrid models combining ML with physical simulations, and constraint-based learning, and highlight their roles in solving complex problems such as material property prediction, microstructure analysis, and phase stability. We also examine how data biases in training datasets can propagate errors and inequities in model outputs, and discuss the ethical values underpinning the use of AI in scientific research, including transparency, accountability, and societal impact. The review underscores the potential of these methods to accelerate innovation in materials science while emphasizing the need for rigorous validation and interdisciplinary collaboration. By synthesizing current advancements, this article aims to provide a foundational understanding for researchers and practitioners, paving the way for future developments in this interdisciplinary field.

Autonomous and semi-autonomous laboratories represent a transformative paradigm in materials science, integrating artificial intelligence, robotics, and high-throughput experimentation to accelerate discovery and optimization processes. This review examines the conceptual foundations of these systems, including closed-loop optimization, machine learning algorithms, and modular hardware architectures. We explore their applications in areas such as alloy development, perovskite synthesis, and nanoparticle engineering, highlighting successes that have reduced discovery timelines from years to days. However, we also critically assess associated risks, including data quality issues, algorithmic biases, ethical concerns in resource allocation, and potential safety hazards from unsupervised operations. Drawing on recent advances, we propose balanced implementation strategies that maximize innovation while mitigating risks. The review underscores the need for interdisciplinary collaboration to realize the full potential of these technologies in addressing global materials challenges.

The ambiguous usage of the term “surprise” in AI-guided discovery literature represents a significant conceptual barrier in artificial intelligence for materials science. Surprise is variously treated as a statistical anomaly flagged by machine learning models, a human psychological state of unexpectedness that prompts belief revision, an information-theoretic measure of divergence between prior and posterior beliefs, or an unexpected breakthrough that leads to a genuine scientific advance. This lack of precision confuses researchers, fragments the literature, and impedes the systematic design of AI systems capable of deliberately cultivating the forms of unexpectedness that drive materials innovation. This paper proposes a precise typology of scientific surprise consisting of four distinct types—predictive surprise, representational surprise, discovery surprise, and methodological surprise—tailored specifically to the domain of AI-guided discovery in materials science. The key distinctions among these types are articulated along four core dimensions: the source of the surprise (originating in the AI model or in the human scientist), the trigger (prediction error, out-of-distribution data, contradiction with existing theory, or unexpected patterns in the inquiry process itself), the experiencer (primarily the model or the scientist), and the epistemic consequences that follow (model retraining, expansion of representational capacity, theory revision, or redesign of search and measurement strategies). By furnishing this conceptual framework, the paper offers clear implications for designing AI systems that can report, distinguish, and cultivate productive forms of surprise, thereby transforming AI from a passive predictor into an active partner in the discovery process and enabling more effective, targeted responses to different kinds of unexpectedness in materials science.

This review article examines the literature on scientific explanation in AI-driven materials science, focusing on the conceptual foundations and evaluative criteria that distinguish genuine scientific explanation from the predictive and interpretive outputs commonly produced by machine learning models in the field. The methodology involved a systematic search across major databases and targeted journals using predefined strings related to scientific explanation, explainable AI (XAI), and interpretability in materials contexts, resulting in the inclusion of 30 peer-reviewed publications from 2017 to 2024 that directly address the intersection of philosophical theories of explanation and practical AI applications in materials discovery and property prediction. Philosophical theories of explanation, including the deductive-nomological model of Hempel and Oppenheim, the causal-mechanical account advanced by Salmon, unificationist approaches that emphasize the integration of disparate phenomena, and pragmatic frameworks that treat explanations as context-dependent answers to why-questions, provide essential benchmarks against which current materials AI practices can be assessed. In current materials AI literature, explanation is frequently conflated with prediction or post-hoc interpretability techniques such as feature importance scores and attention visualizations, as seen in comprehensive surveys of machine learning for molecular and materials science and recent advances in solid-state applications. Yet, these approaches often remain correlational rather than mechanistically grounded. XAI methods applied to materials problems, including SHAP-based feature attribution, attention mechanisms in graph neural networks, surrogate modeling, and counterfactual generation, offer valuable local insights but fall short of meeting the standards of scientific explanation due to their inherent limitations in capturing causality, multi-scale mechanisms, and physical plausibility. Ultimately, this review articulates adapted criteria for scientific explanation tailored to materials science’s multi-scale and emergent challenges and proposes actionable recommendations to bridge the gap between XAI outputs and robust explanatory accounts, urging the community to prioritize mechanistic understanding over mere predictive accuracy to advance trustworthy and insightful AI-driven discovery.

Artificial intelligence is rapidly moving beyond its early role as a pattern-recognition and predictive-modelling tool in materials science. What began as an acceleration strategy for screening known datasets is now becoming a broader transformation of how materials hypotheses are generated, tested, and refined. The central problem is that this transformation is often described in fragments: predictive models in one literature, generative design in another, physics-informed learning in another, and autonomous laboratories in yet another. A unified conceptual synthesis is needed to explain how these streams collectively move AI from passive assistant to active scientific collaborator. This integrative review traces the evolution of AI in materials science from 2017 to 2026. It frames the field through the idea of the AI co-scientist: an intelligent system that can recognise patterns, propose candidates, incorporate physical constraints, select experiments, and learn from feedback. The review integrates 31 peer-reviewed articles spanning materials informatics, machine learning, generative AI, inverse design, physics-informed modelling, active learning, autonomous experimentation, and self-driving laboratories. It does not present new empirical data, meta-analysis, or bibliometric mapping. The synthesis identifies four major evolutionary stages: pattern recognition, generative design, physics-integrated AI, and autonomous experimentation. These stages are not isolated phases but mutually reinforcing capabilities that increasingly connect computation, synthesis, characterisation, and human judgement. The review concludes that AI is becoming a genuine partner in materials discovery, but this transition depends on trustworthy data infrastructure, interpretable models, robust experimental integration, and new norms for human–AI collaboration. The co-scientist paradigm offers a forward-looking framework for understanding how materials science may be reorganised around closed-loop intelligence.

This review examines the problem of scientific consensus formation in AI-driven materials science by systematically analyzing conceptual approaches from philosophy and sociology of science alongside empirical developments in computational materials research, drawing exclusively on 31 peer-reviewed publications from 2017–2026 identified through targeted searches in Web of Science, Scopus, arXiv, and PhilPapers using terms such as “scientific consensus” AI materials, “consensus formation” machine learning science, “disagreement” materials AI, “benchmark” consensus materials informatics, “epistemic consensus” AI science, “paradigm” materials AI, “scientific disagreement” computational science, and “consensus mechanism” AI research, with inclusion criteria limited to papers addressing epistemology, disagreement, uncertainty, benchmarks, or paradigm dynamics in data-driven disciplines and exclusion of purely technical performance reports. Consensus concepts are traced from logical-positivist agreement on theories through Kuhnian paradigms and Mertonian social processes to Bayesian convergence and pragmatic problem-solving necessities, revealing how each framework illuminates different facets of knowledge coordination in materials science. AI’s impact on consensus formation operates through six distinct mechanisms—accelerated hypothesis validation, model disagreement, benchmark-driven focal points, opacity-induced dissent, data-driven convergence, and authority shifts—both facilitating rapid agreement on material properties and simultaneously generating new forms of epistemic fragmentation. These dynamics create profound tensions and paradoxes, including the trade-off between speed and deliberation, convergence versus diversity, predictive agreement versus explanatory understanding, local versus global consensus, and human versus AI authority, while exposing critical gaps such as the absence of a dedicated theory for AI-mediated consensus, the scarcity of empirical studies tracking real-time consensus processes in materials AI communities, and unresolved questions about managing productive disagreement. Recommendations are offered for researchers, journals, and the broader community to distinguish model agreement from scientific consensus, institutionalize empirical consensus studies, preserve productive dissent, and develop governance protocols that harness AI’s epistemic power without sacrificing critical scrutiny, thereby guiding the field toward more reflexive and robust knowledge production in the age of AI-augmented materials discovery.

The field of materials science has undergone a transformative shift with the integration of computational and data-driven approaches, particularly through representation learning techniques that enable efficient handling of complex materials data. This review synthesizes recent advancements in architectures for representation learning, encompassing graph neural networks, attention-based models, and physics-inspired embeddings, which facilitate the extraction of meaningful features from diverse data modalities such as atomic structures, stoichiometries, and spectroscopic data. By bridging traditional computational methods with machine learning, these representations have accelerated property prediction, inverse design, and materials discovery applications, addressing challenges in high-dimensional spaces and sparse datasets. The scope of this narrative review covers the evolution from basic informatics to sophisticated multimodal integrations, highlighting how data ecosystems and learning frameworks contribute to autonomous discovery pipelines. A systems-level perspective is adopted to integrate cross-study insights, revealing synergies between representation learning and closed-loop systems that couple simulations with experiments. Looking ahead, the review posits that continued refinement of these architectures will drive scalable, AI-guided materials engineering, fostering innovations in energy, electronics, and structural materials while emphasizing the need for robust, interpretable models in real-world applications.

The advent of foundation models, large-scale pre-trained architectures adapted from natural language processing paradigms, has permeated computational materials science, promising accelerated discovery through data-driven inference. In materials engineering, these models leverage multimodal datasets encompassing atomic structures, properties, and simulations to enable representation learning across scales. However, inherent conceptual limits arise from the interplay between materials' physical hierarchies—spanning quantum to macroscopic levels—and the inductive biases embedded in pretraining strategies. This manuscript synthesizes recent advancements in machine learning architectures, such as graph neural networks and multimodal integration, within materials informatics ecosystems. It identifies epistemic boundaries where foundation models falter in capturing causality, uncertainty, and domain-specific invariances, potentially leading to misaligned discovery pipelines. To address these, we introduce the Matter Pretraining Boundary Framework (MPBF), a conceptual architecture that delineates layers of data assimilation, representational abstraction, and inference steering to mitigate limits in autonomous materials design. Implications extend to high-throughput computation, inverse design, and simulation-experiment coupling, fostering more robust computational workflows in materials engineering. By interpreting these limits through systems-level dynamics, the framework guides infrastructure trade-offs, enhancing the reliability of data-driven paradigms without empirical validation.

The convergence of large-scale machine learning with materials engineering is reshaping how new materials are conceived, predicted, and realized. Foundation models—pre-trained architectures that learn generalizable, multimodal representations from expansive datasets—are emerging as the computational backbone for next-generation discovery pipelines. This narrative review synthesizes the computational and data-driven ecosystems that have enabled their rise, drawing on advances in materials informatics, graph-based representation learning, and autonomous experimentation. We trace the progression from early machine learning applications in property prediction to scalable graph neural networks that capture atomic-scale interactions with unprecedented fidelity. High-throughput computation and multimodal data integration have created the knowledge bases necessary for training models that generalize across chemical spaces. Central to this evolution are closed-loop systems, where foundation-like models orchestrate active learning, uncertainty-aware selection, and seamless simulation–experiment feedback. Through an original integrative analysis, we identify recurring architectural principles—such as hierarchical graph convolutions, contrastive pre-training, and multi-task optimization—and training paradigms that balance exploration with exploitation in vast design spaces. These elements collectively address longstanding bottlenecks in inverse design, property optimization, and length-scale bridging. Positioned at the interface of computational infrastructure and autonomous discovery, this review provides a systems-level perspective on how foundation models are poised to compress the materials innovation timeline from decades to months, while maintaining rigorous physical grounding.