Towards Agentic Intelligence for Materials Science

2 Recent Progress of AI and LLMs

• AI4Materials research is traced through predictive models, foundation models, post-training, and agentic systems with tool use and long-horizon rewards.Pretrained foundation models initialize the materials-science agentic pipeline, so pretraining choices influence downstream capabilities and can be adapted using reinforcement-learning signals from real-world experiments.Source: §2 · pp. 5–6

2.1 Deep Learning Revolution

• Classical machine learning enabled materials property prediction, pattern recognition, phase identification, and systematic chemical-space exploration, but lacked multimodal representation learning.Supervised methods established structure–property relationships, unsupervised methods extracted patterns from high-dimensional data, and reinforcement learning optimized sequential decisions over design spaces.Source: §2.1 · pp. 6, 8

2.2 Transformer and Foundation Models

• Transformers and foundation models expanded scientific language, reasoning, and multimodal capabilities, while scaling introduced risks in bias, hallucination, calibration, controllability, and verification.Materials applications remain constrained by scarce high-quality data, limited robust 3D scientific understanding, and the need for domain-specific calibration and evaluation.Source: §2.2 · pp. 8–10

2.3 LLMs and Agents

• LLM agents move beyond passive prediction by combining reasoning, persistent memory, planning, and external tools to execute multi-step materials research and closed-loop discovery.Agents can read literature, generate hypotheses, plan experiments, design syntheses, retrieve resources, and coordinate actions, but reliable grounding, stability and synthesizability constraints, tool reliability, feedback design, safety, and closed-loop validation remain limiting factors.Source: §2.3 · pp. 10–11

2.4 Pipeline-Centric Perspective

• The pipeline-centric perspective treats pretraining as a controllable module whose data, objectives, and safety constraints interact with domain adaptation and agentic feedback loops.This framing emphasizes system-level integration of feedback, coordination, and objective alignment rather than isolated model capabilities.Source: §2.3 · p. 12, §2.4 · p. 12
• Broad web-scale pre-training may be unnecessary or suboptimal for specialized science, while poisoned or misleading data can induce harmful behaviors despite unaffected benchmarks.General-purpose agents must also be tailored to materials workflows, safety regulations, and experimental constraints.Source: §2.4 · pp. 12–13
• Pre-training should become an adaptive feedback-driven loop in which materials-science outcomes and safety jointly determine data amplification, attenuation, removal, and domain-corpus expansion.Long-horizon rewards from simulation and experiments can propagate backward to revise corpus composition, sampling strategies, and alignment objectives when inherited behaviors hinder discovery.Source: §2.4 · p. 13
• Validated discovery outcomes should serve as the primary reward signal, linking upstream training choices to experimentally validated materials rather than surrogate benchmarks alone.This reframes better pre-training around discovery-relevant behaviors such as feasible synthesis planning and hazard handling.Source: §2.4 · p. 13
• Influence functions can identify training examples responsible for emergent behaviors, enabling goal-driven data selection that emphasizes discovery capabilities and removes sources of hallucinated mechanisms or impossible synthesis conditions.Approximate influence estimators use scalable inverse-Hessian approximations to recover sparse, interpretable training sequences linked to model behaviors.Source: §2.4 · p. 13
• Pre-training, domain adaptation, and instruction tuning form a continuously reshaped, influence-addressable memory driven by rewards from agentic simulation and experimental closed loops.Influence functions thereby become prospective levers for dynamically realigning training with novel, useful, and safe materials discovery while mitigating open-world data risks.Source: §2.4 · p. 13

• The section examines traditional AI progress across predominantly reactive materials-science tasks to assess its limitations for autonomous materials discovery.It focuses on integrating information and decision-making across the composition-process-structure-property-performance chain.Source: §3 · p. 14
• High performance on isolated tasks such as property prediction or information extraction can misrepresent discovery value when synthesis, safety, and experimental validation are excluded.These benchmarks are treated as potentially misleading proxies for end-to-end discovery outcomes.Source: §3 · p. 14
• Materials-science-tuned LLMs are building blocks for agentic systems, but efficient autonomous discovery requires pipeline-level objective alignment and credit assignment from outcomes to upstream components.Individual module optimization may not translate into better discovery performance without connecting real discovery outcomes to upstream design choices.Source: §3 · p. 14

3.1 Prediction

• LLMs combine symbolic structural inputs and high-dimensional features, improving multimodal prediction and enabling assessment of model trustworthiness.Inputs include formulas, space-group/Wyckoff descriptors, crystallographic text, geometric features, DFT-derived features, and foundation-model embeddings.Source: §3.1 · p. 14

3.1.1 Regression Tasks

• Regression models predict continuous properties, but LLMs still struggle with domain shifts, new targets, and synthesis- or processing-relevant constraints.Transformer representations capture global context, long-range interactions, and symmetries, while improved representations alone do not solve long-horizon planning or feedback coupling.Source: §3.1.1 · pp. 14, 17

3.1.2 Classification Tasks

• Classification supports defect analysis, quality control, and experimental automation, but remains largely reactive unless integrated with planning, control, verification, and corrective feedback.LLMs can combine noisy text, images, and numerical data, while conversational systems contextualize defect predictions and suggest process adjustments.Source: §3.1.2 · pp. 18–19

3.1.3 Advanced Methodologies

• Advanced methodologies combine local and global representations, share information across related properties, and quantify uncertainty for active learning, though distribution shift and negative transfer remain limitations.Uncertainty estimates can guide data collection and out-of-distribution detection, but reliable discovery requires propagating uncertainty through multistep planning and experimentation.Source: §3.1.3 · pp. 19–20

3.2.1 Information Extraction

• LLM-based extraction systems use prompting, fine-tuning, table awareness, verification, and tool use, yet remain largely reactive unless coupled to discovery-level objectives and outcome-grounded evaluation. [240, 241, 242, 243, 120, 218, 244, 245]Earlier approaches provided reproducible extraction pipelines but were constrained by fixed grammars and limited adaptability, while MatSciBERT178 and MechBERT239 improved robustness and accuracy across diverse publication sources.Source: §3.2.1 · p. 21

3.2.2 Knowledge Graph

• Knowledge graphs integrate extracted entities, parameters, and outcomes to support reasoning across composition–structure–property relationships, including MatKG and terminology-based networks135,136.Hybrid methods combine GNNs, ontology models, and LLMs, while semi-supervised pipelines reconstruct synthesis–processing–property relations and phase diagrams from literature.Source: §3.2.2 · p. 21

3.2.3 Database Automation and Annotation

• Automated corpora, benchmarks, modular curation pipelines, multimodal extraction, and instrument/database integration enable scalable database construction, but proxy benchmarks may diverge from discovery outcomes.Knowledge-graph coverage and correctness remain bounded by extraction noise and ontology choices, allowing errors to propagate into downstream reasoning and planning248.Source: §3.2.3 · pp. 21–22, §3.2.2 · p. 21

3.3 Generation

• Generative AI supports proactive materials innovation through structure generation, inverse design for desired properties, and synthesis-route generation.These tasks expand materials R&D beyond data analysis into designing novel materials and synthesis methods.Source: §3.3 · p. 22

3.3.1 Structure Generation

• Structure-generation systems increasingly shift from generative sampling toward search-with-verification, where candidates are iteratively generated, evaluated, and improved.Generation may use diverse representations and model families, but discovery depends more on coupling generation to evaluation in the loop; many evaluations still rely on simulated or surrogate criteria.Source: §3.3.1 · p. 22

3.3.2 Inverse Design

• Inverse-design frameworks combine generative models, property predictors, or reinforcement learning to iteratively produce materials targeting specified properties.Their effectiveness depends on surrogate property models and feasibility constraints, which motivate closed-loop verification because proxy optimization may not translate to discovery outcomes.Source: §3.3.2 · p. 23

3.3.3 Synthesis Route Generation

• Synthesis-route generation addresses precursors, reaction paths, and conditions, but feasibility and process constraints remain the bottleneck separating generated candidates from experimentally validated materials.Graph methods, literature-trained predictors, and precursor-similarity models support pathway and recipe generation.Source: §3.3.3 · p. 23
• Autonomous synthesis platforms close the loop by using experimental outcomes to update models, trigger replanning, and provide discovery-level feedback for end-to-end credit assignment.This evolution combines AI planning, robotics, active learning, deep learning, reinforcement learning, and uncertainty-aware search.Source: §3.3.3 · p. 23

3.4.1 Materials Discovery Process Optimization

• Bayesian closed-loop optimization identified low-temperature, high-conductivity palladium thin-film formulations that were later validated in scalable manufacturing, while ARES increased carbon-nanotube growth rate eightfold.BAX variants outperformed strong baselines in locating target materials efficiently, but only in low-dimensional settings.Source: §3.4.1 · p. 24

3.4.2 Simulation and AI-based Verification

• Multi-fidelity Bayesian optimization, GNN–DFT workflows, deep reinforcement learning, and ML interatomic potentials support adaptive verification and reduce simulation costs for longer-horizon discovery loops.Verification combines approximate and accurate sources, while CAMP improves local-environment and higher-order-interaction modeling through physically motivated atomic moment tensors.Source: §3.4.2 · p. 24

3.4.3 Agent-based Closed-loop Labs

• Agent-based closed-loop labs combine automated experimentation with Bayesian search, active learning, machine learning, and high-throughput characterization to discover and optimize photocatalysts, quantum dots, perovskites, and inorganic materials.Examples include the mobile robotic chemist, Artificial Chemist, CAMEO, A-Lab, and an agent-based platform for perovskite solid solutions.Source: §3.4.3 · pp. 24–25
• LLM-centered systems extend autonomy by generating instrument-control scripts, analyzing outcomes, and improving subsequent measurements, while human feedback can dynamically refine objectives during experiments.One implementation controlled Python and LabVIEW instruments from natural-language prompts with minimal human involvement; BOARS located human-desired optima faster than manual optimization.Source: §3.4.3 · p. 25

3.5 Data and Knowledge

• Scarce, heterogeneous materials data and costly real-world evaluation remain central challenges for end-to-end agentic LLM systems.These difficulties are especially pronounced in never-ending experimental environments.Source: §3.5 · pp. 25–26

3.5.1 Data scarcity and heterogeneity

• LeMatTraj285 combines nearly 120 million high-quality atomic configurations from a primary repository into a single dataset, providing the high-quality, standardized data needed to develop strong machine-learning interatomic models.MatMiner281 and Robocrystallographer282 provide tools for turning different types of materials into standardized numerical descriptors.Source: §3.5.1 · p. 26
• Physics-informed and generative augmentation methods mitigate scarcity and mismatch, including a perturbation technique that improved unrelaxed structure prediction accuracy by 66%.Other approaches incorporate experimental errors, elasticity-based strain responses, GAN-generated spectra, and conditional generative data for extreme scarcity.Source: §3.5.1 · p. 26
• Multi-fidelity learning fuses data with varying precision, including sparse high-fidelity data, while MD-HIT289 filters overlapping samples to prevent performance overestimation and enable more realistic evaluation.MODNet163 learned differences between high- and low-quality values and significantly improved results versus learning solely from high-quality experimental data.Source: §3.5.1 · pp. 26–27

3.5.2 Knowledge integration

• Knowledge integration improves model accuracy, interpretability, and generalization by embedding physical laws, structural relationships, ontologies, and knowledge graphs.PINNs encode governing equations, graph models represent atoms and bonds, and element-centric knowledge graphs unify structural and semantic chemical information.Source: §3.5.2 · pp. 27–28
• Materials-specific pretrained models such as MatSciBERT178 and MaterialBERT295 outperform vanilla BERT on identifying material-related properties and phrases, while UniMat177 combines multiple modalities.Specialized models also target batteries, optical materials, polymers, and metallurgy, and MELT300 transfers broad-domain knowledge through semantic knowledge graphs.Source: §3.5.2 · p. 28

3.5.3 Multimodality

• Agentic LLMs must jointly interpret multimodal pipeline outputs because materials research combines text, structural graphs, microscopy images, spectroscopic signals, and other representations.Current resources remain difficult to reproduce across modalities because common metadata and ontologies are lacking; microscopy and multiphase XRD also retain important limitations.Source: §3.5.3 · p. 29
• mCLM312 uses specialized encoders and function- and synthesis-aware knowledge to generate synthesizable molecular structures from natural-language functional requirements.This addresses models that generate targets without upfront synthesis considerations; the closed loop returns structures with the best synthesizable functional modules.Source: §3.5.3 · p. 30
• Unified multimodal models integrate text, crystal graphs, microscopy, spectroscopy, and density-of-states representations for transfer learning, retrieval, and materials property prediction.MultiMat181 jointly encodes four modalities, MatBind322 aligns three modalities through contrastive learning, and MATMMFuse182 combines crystal graphs with SciBERT embeddings for formation-energy and electrical-characteristic prediction.Source: §3.5.3 · pp. 30–31
• Multimodal robustness improves when information is complementary or missing, while language-model and hybrid architectures extend sequential modeling and atomic-text integration to crystallography.MatMCL323 maintains prediction accuracy when one or more modalities are unavailable; DiffractGPT306 analyzes diffraction spectra, and M3GNet179-LLM hybrids link atomic representations with textual context.Source: §3.5.3 · p. 30
• Current multimodal systems often capture correlations rather than mechanisms, incur substantial data-engineering costs, and lack consensus benchmarks for multi-step reasoning, cross-modal generalization, and closed-loop discovery.Existing benchmarks emphasize recognition and alignment over scientific reasoning; future work calls for physical invariances, interoperable ontologies, standardized datasets, and evaluation by experiment selection, validation throughput, and constraint satisfaction.Source: §3.5.3 · pp. 30–31

3.6 Explainability

• The survey consolidates explainability into physical validity, faithfulness, and stability, covering sparse models, attention and graph explainers, and physics-informed methods.Physical validity concerns agreement with scientific reasoning and laws; faithfulness concerns reflecting the model’s internal process; stability concerns consistency under similar inputs or perturbations.Source: §3.6 · pp. 31–32

3.6.1 Sparse and Closed-form Models

• Sparse and closed-form models provide human-readable formulas with strong physical validity and faithfulness, but their symbolic forms can be unstable under noisy or heterogeneous data.They typically depend on curated feature sets and may struggle with raw, high-dimensional modalities, motivating complementary neural explainers.Source: §3.6.1 · p. 32

3.6.3 Physics-informed Interpretability

• Because symbolic, neural, and physics-informed approaches have complementary limitations, the survey motivates hybrid pipelines combining explicit physical rules with data-driven learning.Such combinations are intended to support stronger agentic pipelines while preserving interpretable and physically constrained behavior.Source: §3.6.1 · p. 32, §3.6.2 · p. 33, §3.6.3 · p. 33

3.6.2 Attention and Graph Explainers

• Attention and graph explainers expose influential atoms, bonds, compositions, or spectral features, offering plausible instance-level insights while facing unresolved faithfulness and stability problems.Attention weights may not reflect actual decision logic, and small input perturbations or random-seed changes can alter highlighted structures or spectra.Source: §3.6.2 · pp. 32–33

3.6.3 Physics-informed Interpretability

• Physics-informed methods embed physical principles into model design, improving physical validity, faithfulness, stability, and safety by constraining predictions and actions.Symmetry-equivariant architectures and constraints based on conservation or stress–strain relationships act as feasibility guards, although complex or incomplete laws can limit flexibility and generalization.Source: §3.6.3 · p. 33

Pipeline limitations

• Fine-tuned models largely inherit pretrained behaviors because materials-science data are orders of magnitude smaller, making desired behaviors difficult to elicit through limited domain fine-tuning.This motivates jointly considering materials-science fine-tuned LLMs and agentic LLMs for pipeline decision-making.Source: §3.7 · p. 33
• Current AI4MatSci systems target narrowly scoped supervised tasks with sparse, noisy datasets, leaving a substantial gap between benchmark success and real experimental impact.These datasets remain partial proxies for open-ended discovery and are often optimized independently of end-to-end outcomes.Source: §3.7 · p. 34

Pipeline-centric optimization

• Materials-science tasks should function as adjustable operators whose design, data, and evaluation are continuously re-examined against the ultimate discovery reward.Limited high-quality domain data should be allocated to pipeline components that most directly advance experimental outcomes rather than inflate intermediate scores.Source: §3.7 · p. 34
• The field should shift from reactive intermediate benchmarks toward agentic systems that explicitly close the loop with simulation and experiment.Over-optimizing fragile surrogate metrics can produce local gains without faster, safer identification and validation of functional materials.Source: §3.7 · p. 34

Safety and alignment

• Greater pretraining or adaptation plasticity risks eroding fundamental safety and security properties, so scarce noisy supervision must be coupled with general pretraining to preserve broad alignment and robustness.The intended outcome is autonomous, safe materials discovery without sacrificing foundational principles inherited from general-purpose training.Source: §3.7 · p. 34

• Specialized modules for property prediction, extraction, and generation achieve high performance on static benchmarks but remain fragmented and cannot independently form a discovery engine.Source: §4 · p. 34
• The pipeline-centric perspective requires integrating these capabilities into coherent, goal-directed loops that plan, execute, and learn from experiments.Source: §4 · p. 34
• Agentic systems transition materials-science AI from tools that passively handle assigned tasks to active processors capable of closing the discovery loop.Source: §4 · p. 34

• Agentic systems replace passive structure–property prediction with policies that select experiments or compositions to maximize long-term rewards balancing performance, resource cost, and safety.Their closed-loop formulation uses system state, actions, and feedback from experiments or simulations rather than predictive accuracy alone.Source: §4.1 · p. 36
• Non-convex, expensive, and scientifically constrained search spaces make uncertainty quantification, active learning, and surrogate modeling essential for efficient exploration.These components distinguish fully agentic pipelines from simpler Bayesian optimization loops by coupling richer state representations with guided exploration.Source: §4.1 · p. 36
• Self-driving laboratories close the materials loop through autonomous design, synthesis, evaluation, and model retraining, while hierarchical systems divide discovery among Predictor, Planner, and robotic Verifier agents.Multi-agent architectures communicate through standardized protocols, enabling modular expansion, shared memory, and greater explainability.Source: §4.1 · pp. 36–37
• Safe exploration balances novel compositions against reliable performance using UCB, Thompson sampling, and epistemic uncertainty, while hybrid simulation–experiment loops use DFT or MD as virtual laboratories.Fast models such as DeepMD344 and CHGNet339 can generate large numbers of virtual experiments before physical synthesis, linking computational agents with experimental counterparts.Source: §4.1 · p. 37
• Future scientific multi-agent ecosystems will combine LLM reasoning and documentation with RL planners and simulation engines, but require governance for safety and reproducibility.These AI laboratories are envisioned to support discovery, hypothesis explanation, literature grounding, and experimental justification in natural language while compressing scientific feedback cycles.Source: §4.1 · p. 37

4.2 Scientist AI: Beyond Data Fitting to Scientific Reasoning

• Scientist AI systems aim to formulate hypotheses, orchestrate virtual or physical experiments, and update internal models from evidence rather than only learning structure–property mappings.Existing deep learning models such as CGCNN99 and ALIGNN343 cannot articulate hypotheses, mechanisms, or explanatory frameworks, while robot scientists and self-driving laboratories demonstrate parts of the iterative workflow379,380.Source: §4.2 · pp. 37–38

4.2.1 Hypothesis Generation and Search Over Scientific Spaces

• Bayesian optimization reduces materials-design trials by several orders of magnitude, while generative models expand search toward chemically valid, previously unrecognized structures and constrained inverse-design candidates.BO supports uncertainty-aware exploration–exploitation and closed-loop workflows for alloys, catalysts, and thin films; generative models produce candidates tailored to target properties, stability, and performance constraints.Source: §4.2.1 · p. 38
• LLM-based systems can extract scientific relationships, provide mechanistic rationales, and coordinate reaction or synthesis planning, but they still rarely produce explicit, falsifiable, long-horizon hypotheses grounded in mechanisms and reliable uncertainty.Their limitations include reliance on surrogate predictors, fragile mechanism-to-protocol reasoning, and unresolved factual reliability, calibration, and domain grounding347,350,386,387.Source: §4.2.1 · pp. 38–39

4.2.2 Scientific Critical Thinking

• Scientist AI requires deliberate critical thinking that evaluates competing hypotheses, conflicting evidence, counterfactual designs, and uncertainty under scarce, noisy, costly, and multi-objective scientific feedback.A proposed System 2–style agent would assess prediction plausibility, generate alternative hypotheses, weight claims, anticipate negative results, and use active learning to prioritize uncertain chemical-space regions348,351,388,389,390.Source: §4.2.2 · p. 39

4.2.3 Experiment and Simulation Planning (Decision-Making Under Uncertainty)

• Experiment and simulation planners select informative actions under uncertainty, budgets, costs, and safety constraints, but current systems optimize scalar objectives more often than discriminating mechanisms or updating structured hypotheses.Self-driving laboratories connect planning and automation to experimental outcomes for end-to-end credit assignment349, while LLM agents can coordinate tools and multi-step campaigns but remain vulnerable to hallucinations, poor calibration, and safety violations122,385.Source: §4.2.3 · p. 40
• Scientist-level discovery planning will require a cohesive framework combining uncertainty-aware optimization, explicit hypothesis representations, constraint-aware scheduling, and verifiable automated control.This integration is presented as necessary for comprehensive materials discovery rather than isolated optimization or tool invocation.Source: §4.2.3 · p. 40

4.2.4 Interpretation, Explanation, and Hypothesis Revision

• Attribution and saliency methods connect predictions to structural, compositional, processing, atomic, bonding, or subgraph features and can test mechanistic coherence against new data.However, these explanations typically do not preserve or revise an explicit hypothesis set.Source: §4.2.4 · p. 41
• Symbolic regression and SISSO provide human-readable descriptors or analytical laws that can serve as explicit hypotheses for examination, scrutiny, and modification.These approaches can furnish candidate hypotheses rather than only opaque parameter updates.Source: §4.2.4 · p. 41
• AI Feynman and related equation-discovery frameworks show that systems can derive governing equations and select among symbolic hypotheses instead of only updating neural weights.Active-learning workflows can also transition between functional forms as new data emerges, although demonstrated settings remain idealized.Source: §4.2.4 · p. 41
• Current explainability remains limited under distribution shift or dataset bias, while symbolic techniques are restricted to low-dimensional settings and require substantial expert specification.LLM- and agent-based systems can generate rationales or alternative mechanisms, but these are rarely tied to formal hypothesis sets or evaluated across iterative data collection.Source: §4.2.4 · p. 41
• Scientist-level interpretation demands explicit hypothesis tracking, uncertainty-aware selection among competing mechanistic models, and benchmarks measuring revision within a closed-loop discovery process.The goal is to assess whether AI systems revise and refine hypotheses, not merely satisfy diagnostic explainability criteria.Source: §4.2.4 · p. 41

Closed-loop collaboration

• Closed-loop laboratories can connect AI reasoning, molecule proposal, synthesis, testing, and real-world feedback for continual improvement alongside human scientists.This addresses the limitation that existing molecular AI systems often remain static learners unable to adapt to new experimental evidence.Source: §4.3 · p. 41

Human–AI workflow integration

• Interpretable GNNs and LLM interfaces support two-way collaboration by explaining structure–property relationships, contextualizing results, retrieving knowledge, and guiding experiments across the scientific lifecycle.Human scientists assess physical plausibility and refine hypotheses, while AI identifies high-dimensional correlations and integrates textual, numerical, and structural information.Source: §4.3 · p. 42

Human oversight and interfaces

• As systems such as ChemOS automate low-level planning and execution, humans should emphasize management, ethical oversight, supervision, conceptual innovation, and assessment of structured AI hypotheses.Interpretable feedback loops and interactive interfaces help users understand data, model sensitivity, and why an AI system will act.Source: §4.3 · p. 42

Collaboration networks

• Human–AI collaboration networks assign humans goal-setting, execution, and validation while specialized agents generate and refine hypotheses through shared interfaces resembling research teams.This framing treats agents as complementary team members with distinct roles coordinated through communication and planning.Source: §4.3 · p. 43

Co-creative science

• Co-creative workflows turn hypothesis–experiment–analysis into AI-mediated feedback loops and organize experimental data into shared representations that improve interpretability and reproducibility.These representations create an epistemic language understandable to both humans and machines.Source: §4.3 · p. 43

Trustworthy collaboration

• Trustworthy collaboration requires critic agents, evaluation of generalizability and fairness, transparent reporting, data-source tracking, documentation, monitoring, and safeguards for attribution, authorship, bias, and accountability.Auto-Labs360 uses a critic agent to check safety and feasibility, while GIFTERS evaluates trustworthiness dimensions including interpretability, fairness, transparency, explainability, robustness, and stability.Source: §4.3 · p. 43

• Agentic systems are better suited to materials discovery because hypotheses must be tested through simulations and preferably experiments, making experience-driven learning difficult to avoid.The survey frames interaction with the environment as a concrete demand of materials science rather than resolving the broader debate over reinforcement learning.Source: §4.4 · pp. 43–44
• End-to-end systems should use backward credit assignment to link discovery successes and failures to corpus selection, objectives, model adaptation, and instruction-tuning criteria.Signals such as failed synthesis, inaccurate predictions, and breached physical constraints could guide which pretraining data are amplified, reduced, or removed, enabling longer-term improvement.Source: §4.4 · p. 44
• Most existing agentic systems operate in simulations or synthetic benchmarks rather than real experiments, leaving them exposed to equipment limits, measurement noise, synthesis variability, and irreducible chemical complexity.Simulation-trained agents may identify materials that are stable in silico but impractical or too expensive to produce.Source: §4.4 · p. 44
• Robotic laboratories, experimental feedback, and physical synthesis and characterization should close the simulation-to-reality gap, while upstream pretrained foundations provide priors that experimentation refines toward robust, safety-aware autonomous discovery.The proposed systems must connect affordable experimentation with the messy realities of physical materials workflows.Source: §4.4 · p. 44

Pipeline-centric framework

• The survey unifies corpus curation, pre-training, adaptation, instruction tuning, and agentic experimentation around end-to-end materials-discovery objectives.It integrates machine-learning and materials-science perspectives while organizing reactive tasks, cross-cutting issues, and Scientist-AI workflows within one conceptual framework.Source: §5 · p. 44

Credit assignment

• Experimental reward signals should reshape upstream memory, data, objectives, and weights through credit assignment rather than optimizing proxy benchmarks.Influence functions can trace pre-training and fine-tuning data contributions to novel-material discovery, while the whole pipeline remains tunable against the final reinforcement-learning reward.Source: §5 · pp. 44–46

Memory and learning

• Pre-training, domain adaptation, and lifelong reinforcement learning form unified memory, but reward hacking can make that memory overfit benchmark artifacts instead of real experimental performance.The proposed response is to update shared memory from outcome-level rewards across the pipeline, including simulation, experiments, and upstream training.Source: §5 · pp. 45–46

Virtual–real transfer

• Virtual-world pretraining followed by real-world continual training is practical, but simulator knowledge limits generalization and costly, inaccessible laboratory automation restricts real-world feedback.The survey proposes parallel memory and search across virtual and real environments, while emphasizing that autonomous laboratories remain expensive, slow to deploy, and resource-intensive to adapt.Source: §5 · p. 46

Evaluation and testbeds

• Fixed prediction benchmarks miss iterative and serendipitous discovery, motivating a modular dynamic AI4MatSci testbed that co-evolves agents, virtual worlds, and upstream components from real-world rewards.The proposed architecture can use shared agents or agent populations with memory across simulation, experiments, and pre- and mid-training, while remaining updateable with real-world data.Source: §5 · p. 46

Safety and controllability

• Materials-discovery agents must be trustworthy and safe because autonomous workflows can accelerate beneficial research while lowering barriers to hazardous materials and synthesis routes.The conclusion recommends dynamically regulating agent plasticity, decoupling agents and environments, and placing learning or optimization pressure on data, retrieval, tool use, experiment proposals, and protocols.Source: §5 · pp. 46–47