PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Agentic Intelligence for Materials Science

Huan Zhang, Yizhan Li, Wenhao Huang, Ziyu Hou, Yu Song, Xuye Liu, Farshid Effaty, Jinya Jiang, Sifan Wu, Qianggang Ding, Izumi Takahara, Leonard R. MacGillivray, Teruyasu Mizoguchi, Tianshu Yu, Lizi Liao, Yuyu Luo, Yu Rong, Jia Li, Ying Diao, Heng Ji, Bang Liu

arXiv:2602.00169v2cond-mat.mtrl-scics.AI

TL;DR

Materials-science AI remains fragmented across task-specific models, while autonomous discovery requires systems that plan, act, and learn across closed-loop workflows. This survey presents a pipeline-centric framework spanning data and model development through agentic simulation and experimentation, arguing that upstream choices should be aligned with experimentally validated discovery outcomes. It concludes with a roadmap toward unified, adaptive, trustworthy, and safety-aware agents for novel and useful materials.

  • Problem

    Existing materials-science AI remains limited by task isolation, persistent data scarcity, and incomplete evaluation of multi-step reasoning and closed-loop decision-making.

  • Method

    The survey analyzes AI and materials workflows through a pipeline-centric lens linking pre-training, adaptation, alignment, uncertainty, tools, memory, and agentic closed loops.

  • Results

    The survey proposes evaluating training and system choices through downstream discovery outcomes, including experimentally validated materials, rather than proxy benchmarks alone.

  • Takeaways & Limitations

    A unified pipeline can continuously reshape models and upstream choices toward discovering novel, useful, and safe materials.

  • Takeaways & Limitations

    Current multimodal benchmarks only partially evaluate multi-step scientific reasoning and closed-loop decision-making, while some agentic systems cannot learn beyond simulator knowledge.

Abstract

from arXiv · show

The convergence of artificial intelligence and materials science presents a transformative opportunity, but achieving true acceleration in discovery requires moving beyond task-isolated, fine-tuned models toward agentic systems that plan, act, and learn across the full discovery loop. This survey advances a unique pipeline-centric view that spans from corpus curation and pretraining, through domain adaptation and instruction tuning, to goal-conditioned agents interfacing with simulation and experimental platforms. Unlike prior reviews, we treat the entire process as an end-to-end system to be optimized for tangible discovery outcomes rather than proxy benchmarks. This perspective allows us to trace how upstream design choices-such as data curation and training objectives-can be aligned with downstream experimental success through effective credit assignment. To bridge communities and establish a shared frame of reference, we first present an integrated lens that aligns terminology, evaluation, and workflow stages across AI and materials science. We then analyze the field through two focused lenses: From the AI perspective, the survey details LLM strengths in pattern recognition, predictive analytics, and natural language processing for literature mining, materials characterization, and property prediction; from the materials science perspective, it highlights applications in materials design, process optimization, and the acceleration of computational workflows via integration with external tools (e.g., DFT, robotic labs). Finally, we contrast passive, reactive approaches with agentic design, cataloging current contributions while motivating systems that pursue long-horizon goals with autonomy, memory, and tool use. This survey charts a practical roadmap towards autonomous, safety-aware LLM agents aimed at discovering novel and useful materials.

2 Recent Progress of AI and LLMs

AI for materials science has progressed from task-specific predictive models toward foundation models and agentic systems that use tools, memory, feedback, and long-horizon rewards. A pipeline-centric view aligns upstream training choices with downstream discovery outcomes rather than proxy benchmarks alone.

  • Evolution of AI: AI research evolved through predictive models, foundation models, post-training for controllability, and agentic systems with tool use and long-horizon rewards.These stages mirror the materials-science discovery pipeline.
  • Classical and deep learning: Early materials applications used feature engineering, supervised and unsupervised learning, and reinforcement learning for property prediction, pattern recognition, and chemical-space exploration.These methods established foundations for later deep-learning architectures.
  • Foundation models: Transformers and foundation models expanded representation learning by modeling long-range dependencies and supporting broad language, reasoning, and interaction capabilities.Attention-based architectures also enabled parallel computation and multi-head representation of chemical and scientific information.
  • Agents: Recent agents extend AI beyond isolated prediction by sustaining multi-step analysis, using tools, coordinating skills, and interacting with workflows.Long-term performance remains influenced by tool reliability, feedback design, and safety constraints.
  • Pipeline-centric perspective: Training choices should be evaluated by discovery-relevant behaviors and experimentally validated materials rather than generic language metrics or surrogate benchmarks alone.The proposed evaluation loop connects optimization choices to real experimental reward signals.
  • Pipeline-centric perspective: The survey proposes treating pre-training, domain adaptation, and instruction tuning as an influence-addressable memory reshaped by rewards from simulation and experimental closed loops.This framing supports revising data selection and objectives according to downstream discovery and safety outcomes.

3 Reactive Tasks in Materials Science from an AI Perspective

Traditional AI for materials science has advanced information and decision-making across the composition–process–structure–property–performance chain, but existing systems remain insufficient for autonomous discovery.

  • Motivation: The section asks why existing AI materials-science systems do not yet meet the needs of autonomous materials discovery.The question concerns integrating information and decision-making across the composition–process–structure–property–performance chain.

3.1 Prediction

Materials prediction spans multimodal regression and classification tasks, with LLMs improving global-context modeling while remaining limited in adaptation, uncertainty handling, and closed-loop planning. These methods support screening and early inverse design but require downstream verification.

  • General prediction: LLM materials models process symbolic structures and high-dimensional features, supporting multimodal property prediction and trustworthiness assessment.Inputs include formulas, crystallographic descriptors, geometric features, DFT-derived features, and model embeddings.
  • Regression: Regression models predict continuous properties including stability, stiffness, electrical characteristics, thermal characteristics, and formation energy.Formation energy is used to assess crystalline stability in high-throughput screening.
  • Limitations: LLM prediction systems still struggle with domain shift, new target properties, and synthesis- and processing-relevant constraints.Improved representations alone do not resolve long-horizon planning, feedback coupling, or decision-level constraints in closed-loop discovery.
  • Representations: Transformer representations capture global crystal context and long-range interactions more effectively than locally oriented graph representations in some prediction settings.CrystalTransformer combines chemical and structural feature matrices before multi-head self-attention processing.
  • Inverse design: Retrieval-augmented general LLMs can improve prediction accuracy when domain knowledge sources are curated and verifiable, supporting early inverse design of candidate compositions.Proposed compositions remain subject to surrogate fidelity and downstream verification.
  • Advanced methodologies: Uncertainty quantification supports trustworthy AI, active-learning decisions, out-of-distribution detection, and targeted data collection in closed-loop discovery.Aleatoric uncertainty reflects data noise, whereas epistemic uncertainty reflects limited data coverage and model misspecification.

3.2 Mining

Materials-science information extraction converts experimental and performance descriptions into structured representations that support downstream datasets, benchmarks, knowledge graphs, and reasoning. The field has shifted from rigid rules toward learned and hybrid semantic pipelines, while extraction noise and benchmark drift remain constraints.

  • Extraction foundations: Information extraction transforms experimental descriptions and performance data into structured formats suitable for data-driven materials research.Extraction determines what becomes trainable in downstream datasets, benchmarks, and objectives.
  • Rule-based methods: Early rule-based systems produced databases of synthesis parameters, battery compositions, thermoelectric properties, and mechanical data but had limited adaptability.Fixed grammars constrained performance across variations in scientific writing.
  • Learned extraction: Statistical, representation-learning, hybrid, and semi-supervised methods capture latent relations and identify synthesis–processing–property connections.Hybrid approaches combine GNNs, ontology models, and LLMs to extend knowledge graphs and interpret predicted relations.
  • Knowledge representations: Extracted entities and relations feed knowledge graphs such as MatKG and terminology-based networks for reasoning across composition–structure–property space.Machine learning and text mining have also reconstructed magnetic and superconducting phase diagrams from literature.
  • Limitations: Knowledge-graph coverage and correctness remain bounded by extraction noise and ontology choices, allowing errors to propagate into downstream reasoning and planning.Benchmarks can also freeze proxy objectives when their definitions drift from discovery-outcome evaluation.
  • Integrated pipelines: Automated corpora, benchmarks, modular pipelines, multimodal extraction, and instrument-linked workflows broaden extraction toward unified materials representations.These efforts integrate composition, structure, process, and property information from text and tables.

3.3 Generation

Materials generation is expanding from sampling structures toward goal-directed and agentic systems that generate, evaluate, and iteratively refine candidates. Across materials classes, the survey emphasizes coupling generation with verification and closed-loop feedback, especially because surrogate objectives may diverge from synthesis feasibility.

  • LLMs support structure generation, inverse design, and synthesis-route generation, extending materials R&D beyond analysis toward proactive innovation.
  • Autoregressive language models represent crystal structures as token sequences and can generate lattices, compositions, or CIF files directly.
  • Fine-tuned LLaMA-2 models generated stable inorganic crystal structures as text at nearly double the generation speed of diffusion-based CDVAE baselines.
  • MATLLMSEARCH used a pretrained LLaMA-3.1 with evolutionary search and physical evaluation to iteratively search, evaluate, and improve crystal candidates without additional fine-tuning.
  • The field is shifting from generative sampling to search-with-verification, with discovery quality depending more on coupling generation to evaluation than on model family.
  • Surrogate property models and feasibility constraints remain common, while synthesis outcomes can update models and trigger replanning for end-to-end credit assignment.

3.4 Optimization and Verification

Optimization and verification are moving toward closed-loop systems that connect AI planning, automated experiments, and outcome feedback. The survey highlights efficiency gains while stressing that reported evaluations often remain limited in scope or rely on proxy objectives.

  • Materials Discovery Process Optimization: Closed-loop Bayesian optimization mapped conductivity–temperature trade-offs and identified low-temperature, high-conductivity palladium formulations later validated in scalable manufacturing.
  • Materials Discovery Process Optimization: BAX converts user-defined targets into acquisition functions for autonomous targeted discovery, but its benchmark advantage was demonstrated only in low-dimensional settings.
  • Materials Discovery Process Optimization: An autonomous research system achieved an 8-fold increase in carbon-nanotube growth rate through closed-loop optimization.
  • Closed-Loop Laboratories: Closed-loop platforms combine automated synthesis, characterization, active learning, and AI planning so experimental outcomes can update subsequent decisions.
  • LLM-Enabled Autonomy: LLM-based systems can generate instrument-control code, analyze experimental outcomes, and connect experiment design, operation, data collection, and analysis.
  • Human–AI Collaboration: Human-in-the-loop Bayesian workflows let operators refine objectives during experiments, locating human-desired optima faster than manual optimization.

3.5 Data and Knowledge

Materials AI faces scarce, heterogeneous data and costly real-world evaluation, motivating standardized representations, knowledge integration, augmentation, and multi-fidelity learning. These approaches aim to improve data use, generalization, and alignment with synthesis-aware discovery.

  • Data scarcity, heterogeneity, and the cost of real-world evaluation remain central challenges for end-to-end agentic materials systems.
  • MatMiner, Robocrystallographer, OPTIMADE, FAIR metadata, and LeMatTraj address heterogeneity, interoperability, interpretability, and data volume through shared tools and repositories.
  • LeMatTraj assembled nearly 120 million high-quality atomic configurations into a standardized collection for materials modeling.
  • Data Augmentation: Data augmentation ranges from structural perturbations and molecular dynamics to physics-informed generative methods for mitigating scarcity and heterogeneity.
  • Multi-fidelity Learning: Multi-fidelity learning fuses low- and high-accuracy data to improve predictions while reducing computational and human-resource costs.
  • Multi-fidelity Learning: MD-HIT filters redundant samples in multi-source datasets to prevent performance overestimation and enable more realistic evaluation.
  • Knowledge Integration: Knowledge integration uses materials-science knowledge, physical laws, structural features, and knowledge graphs to guide models toward greater accuracy, interpretability, and generalization.
  • Multimodality: Multimodal benchmarks still emphasize recognition and cross-modal alignment more than multi-step reasoning or closed-loop decision-making relevant to autonomous discovery.

3.6 Explainability

Explainability in materials AI is framed around physical validity, faithfulness, and stability, because autonomous systems require mechanistic insight, error diagnosis, and trustworthy decisions. Existing methods offer complementary strengths but remain limited by instability, weak faithfulness, or restricted input modalities.

  • Explainability matters for autonomous experiments because explanations can serve as verification artifacts that gate high-stakes actions.
  • The survey consolidates explainability into three domain-relevant axes: physical validity, faithfulness, and stability.
  • Sparse and Closed-form Models: These models typically rely on curated feature sets and may struggle with raw, high-dimensional modalities.
  • Sparse and Closed-form Models: Sparse and closed-form models provide physically meaningful, often faithful descriptors, but their explanations can change under small input or feature-set perturbations.
  • Attention- and Graph-based Explainers: Attention- and graph-based explainers identify influential atoms, bonds, or spectral peaks and provide domain-plausible instance-level insights.
  • Attention- and Graph-based Explainers: Their faithfulness is not guaranteed because attention weights may reflect correlations rather than features causally responsible for predictions, and small perturbations can alter explanations.
  • Hybrid Explainability: Hybrid approaches combine attention with perturbation tests or physics-informed constraints to address explanation faithfulness and stability.

3.7 Pipeline-Centric Perspective

The survey argues that materials-science models should be evaluated as components of an end-to-end discovery pipeline rather than as isolated task solvers. This perspective prioritizes allocating scarce data and optimization effort toward experimental outcomes while preserving foundational safety and robustness.

  • Data and task limitations: Most current AI4MatSci formulations focus on narrowly scoped supervised tasks, such as property prediction, information extraction, structure generation, and local optimization.Fine-tuned models inherit many behaviors from pretraining, making desired behaviors difficult to elicit through limited MatSci-specific fine-tuning alone.
  • Data and task limitations: Materials-science datasets are smaller, less curated, sparse, and noisy compared with web-scale pretraining corpora, limiting their coverage of open-ended discovery.These datasets remain partial proxies, leaving a gap between benchmark success and real experimental impact.
  • Pipeline-centric optimization: Intermediate MatSci tasks should be treated as adjustable operators within a larger discovery pipeline, with design, data, and evaluation continuously examined against the ultimate discovery reward.The survey warns that optimizing fragile surrogates can produce local improvements without faster, safer identification and validation of functional materials.
  • Pipeline-centric optimization: The survey advocates shifting effort from reactive intermediate benchmarks toward agentic systems that explicitly close the loop with simulation and experiment.This allocation is intended to direct limited high-quality data toward pipeline components that most directly advance experimental outcomes.
  • Safety and alignment: Increasing pretraining or adaptation plasticity risks eroding foundational safety and security properties because domain-specific tasks do not cover all normative constraints.The pipeline-centric challenge is to couple scarce, noisy MatSci supervision with general-purpose pretraining while preserving broad alignment and robustness.

4 Agentic Systems for Materials Science

The survey frames agentic materials-science systems as a transition from passive tools and fragmented task modules to active processors that can close the discovery loop. True acceleration requires coherent, goal-directed integration of capabilities for planning, execution, and learning.

  • Research direction: The section asks how existing AI for materials science can bridge the gap to a fully end-to-end autonomous materials-discovery pipeline.Figure 7 provides a taxonomy of recent progress in agentic systems for materials science.
  • From reactive tasks to agents: Existing materials-science AI systems mainly address reactive, isolated tasks such as property prediction, extraction, and generation.Although these modules can perform well on static benchmarks, they do not independently constitute a discovery engine.
  • From reactive tasks to agents: True discovery acceleration requires integrating specialized capabilities into coherent, goal-directed loops that can plan, execute, and learn from experiments.This integration marks the transition from tools that passively handle assigned tasks to active processors capable of closing the discovery loop.

4.1 Contemporary Agentic Systems in Materials Discovery

Contemporary agentic materials-discovery systems combine decision-making, uncertainty-aware exploration, and automated experimentation. Their practical operation depends on integrating models with robotic and computational platforms under scientific constraints.

  • Decision-making and optimization: Reinforcement learning selects actions from system state to maximize long-term objectives balancing experimental performance, resource cost, and safety.In materials settings, actions may include synthesis parameters or composition proposals, while state includes prior results and model uncertainty.
  • Decision-making and optimization: Materials-science agentic systems face expensive, non-convex state and action spaces constrained by thermodynamic stability and synthesis feasibility.Uncertainty quantification, active learning, and surrogate modeling guide exploration efficiently under these constraints.
  • Autonomous experimentation: Self-driving laboratories combine robotic synthesis and characterization with AI models that predict outcomes, evaluate uncertainty, and iteratively plan experiments.This learning loop closes the materials loop by linking experiment selection, execution, and updating.
  • Autonomous experimentation: Safe exploration balances novelty against reliable performance by combining exploration–exploitation methods with uncertainty estimates and expected information gain.Hybrid simulation–experiment loops can use DFT or molecular dynamics as virtual laboratories for generating synthetic pretraining data.

4.2 Scientist AI: Beyond Data Fitting to Scientific Reasoning

The survey describes a shift from accurate prediction toward AI systems that formulate hypotheses, coordinate experiments, and revise scientific models. Current methods can expand candidate spaces and support literature-grounded proposals, but cohesive long-horizon scientific reasoning remains incomplete.

  • Scientific reasoning: Scientist AI systems aim to formulate hypotheses, orchestrate virtual or physical experiments, and update internal models from resulting evidence.They maintain possible explanations for how structure, processing, and environment affect behavior rather than only predicting attributes.
  • Search and optimization: Bayesian optimization reduces experimental trials and supports closed-loop workflows by identifying promising regions under uncertainty.It partially implements hypothesis proposal by locating useful areas within existing design-space uncertainty.
  • Hypothesis generation: Generative models expand the effective hypothesis space by proposing chemically valid, previously unrecognized compositions and microstructures tailored to design goals.Inverse-design workflows evaluate candidates against constraints such as stability and performance using high-fidelity simulations or experiments.
  • Hypothesis generation: LLM-based agents can use literature and external tools to suggest reaction pathways, enumerate viable synthetic routes, and recommend experimental parameter sweeps.Retrieval and organized knowledge bases help incorporate explicit priors and limitations into these proposals.
  • Open limitations: Current systems often provide candidate structures without explicit mechanisms, validity regimes, failure modes, or links to falsifiable theories.Long-horizon reasoning from mechanism to candidate design to experimental protocol remains fragile and dependent on limited training distributions or manual prompts.
  • Future direction: A cohesive discovery framework would combine uncertainty-aware optimization, explicit hypothesis representations, constraint-aware scheduling, verifiable platform control, and end-to-end credit assignment.The proposed direction links model updates and planner behavior directly to experimental outcomes rather than relying exclusively on offline proxy metrics.

4.3 Human–AI Collaboration in Scientific Workflows

Human–AI collaboration reframes materials discovery as a co-creative workflow in which specialized agents and humans jointly plan, generate, evaluate, and validate scientific work. Such systems connect speculation, simulation, experimentation, and analysis while requiring interpretability, safety, accountability, and human agency.

  • Existing molecular AI systems remain static learners that fail to adapt to new experimental evidence or improve through real-world feedback.
  • Closed-loop laboratories can iterate through reasoning, proposal, synthesis, testing, and real-world feedback, leveraging agent self-improvement alongside human scientists.
  • Human–AI workflows organize specialized agents and human supervisors through shared interfaces, with humans defining goals and evaluating, executing, and validating results.
  • AI-assisted scientific workflows form feedback loops between symbolic reasoning and data-driven discovery by transforming experimental data into representations understandable to humans and machines.
  • Trustworthy collaboration requires safety mechanisms, interpretability, alignment, transparent reporting, data-source tracking, documentation, monitoring, and attention to attribution, authorship, bias, and accountability.
  • Interpretable models and language systems integrate human intuition with machine reasoning across speculation, simulation, and analysis, while autonomous frameworks preserve human agency through open interfaces.

4.4 Pipeline-Centric Perspective

A pipeline-centric perspective treats agentic materials discovery as an end-to-end system that must connect upstream model choices to downstream experimental outcomes. Current systems remain incomplete because feedback is not assigned backward through the pipeline and many operate only in simulations rather than real laboratories.

  • An agentic system is presented as the more appropriate design choice for materials science because discovery requires hypotheses to be tested through simulations and preferably experiments.
  • Current agentic systems remain incomplete from an end-to-end, pipeline-centric perspective in two critical respects.
  • Backward credit assignment links downstream successes and failures to earlier choices in corpus selection, objective design, and model adaptation.
  • Agentic successes and failures can transmit training signals to pre-training data curation, domain adaptation objectives, and instruction-tuning criteria, enabling influence-driven correction.
  • Without this feedback loop, agentic systems remain separate from their underlying representations and assumptions, making long-term improvement harder.
  • Most existing agentic systems operate in simulations or synthetic benchmarks, which can obscure equipment limits, measurement noise, synthesis variability, and chemistry's irreducibility.
  • Real materials-discovery agents must connect robotic laboratories, experimental feedback, and physical synthesis and characterization, with experimentation refining pretrained priors toward robust, safety-aware discovery.

5 Discussion & Conclusion

The survey proposes a pipeline-centric framework that aligns AI components, agentic systems, environments, and human collaboration with end-to-end materials discovery objectives. It emphasizes real-world outcome signals, shared memory and search, adaptable testbeds, and safety-aware autonomy while identifying simulation, data, laboratory, and benchmarking constraints.

  • Pipeline-centric framework: The survey integrates corpus curation, pre-training, domain adaptation, instruction tuning, and goal-conditioned agentic LLMs into one materials-discovery framework.The framework connects upstream learning stages to agents operating in open-ended experimental environments.
  • End-to-end optimization: Discovery-oriented AI4MatSci requires aligning the full workflow with real experimental reward signals rather than proxy benchmarks.The proposed view supports credit assignment from real-world outcomes back to pre-training, adaptation, and task design.
  • Memory and search: The pipeline treats pre-training, domain-specific fine-tuning, lifelong reinforcement learning, and agent interaction as unified memory and search processes.Pre-training and tuning contribute to memory, while the agent searches through interaction and stores useful information afterward.
  • Deployment constraints: Simulation pretraining can support transfer, but simulator-bound knowledge may leave a performance gap relative to real-world training, which is prohibitively expensive.Real-world laboratory access is constrained by persistent data scarcity, expert dependence, deployment costs, and limited feedback for learning.
  • Future systems: The survey calls for shared-memory agents and modular, dynamically updatable testbeds spanning simulation and experiments while incorporating real-world data.It proposes evolving agents, virtual environments, and upstream components together, with parallel memory and search across virtual and real worlds.
  • Evaluation and safety: Current benchmarks emphasize fixed prediction tasks and isolated computational elements, whereas discovery is iterative, exploratory, and sometimes serendipitous.The survey therefore advocates trainable pipeline components, long-horizon reward signals, and trustworthy agents that pursue novel, useful, and safe materials.
Loading 2602.00169v2…