Paper deep dive
An LLM agent for end-to-end computational materials discovery
Chen Yuntong, Huang Ju, Liu Yu, Zhao Dan, Sun Mingqi, Ju Chentian, Liu Yanbing, Huang Lijiang, Zhao Guobin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/24/2026, 4:19:55 AM
Summary
The paper introduces MAESTRO, an LLM-based multi-agent system designed for end-to-end computational materials discovery, specifically targeting Metal-Organic Frameworks (MOFs). MAESTRO automates the entire pipeline from literature mining and structure validation to property calculation and multi-fidelity screening. In a case study for CO2 capture from wet flue gas, the system identified high-performance MOF candidates (e.g., ZAWZAA, YIWBEN) that originated from unrelated research domains (such as luminescence and magnetism), demonstrating the system's ability to uncover non-obvious materials through cross-domain provenance tracking and rigorous multi-stage screening.
Entities (11)
Relation Signals (9)
MAESTRO → targets → Metal-organic frameworks
confidence 98% · MAESTRO... capable of executing the entire screening pipeline for metal-organic frameworks (MOFs)
MAESTRO → appliedto → CO2 Capture
confidence 95% · We showed a case study for CO2 capture from wet flue gas
YIWBEN → identifiedby → MAESTRO
confidence 95% · YIWBEN, the economic champion, came from a magnetic relaxation study of a Co-MOF.
ZAWZAA → identifiedby → MAESTRO
confidence 95% · ZAWZAA, the performance champion, was first reported in a study of luminescent Schiff-base coordination polymers.
MAESTRO → uses → LLM
confidence 95% · The decision layer is driven by a large language model
MAESTRO → queries → OpenAlex
confidence 90% · The literature agent... retrieved over 400,000 papers from OpenAlex.
MAESTRO → queries → Cambridge Structural Database
confidence 90% · It associated these papers with crystal structures from the Cambridge Structural Database (CSD)
Widom Insertion → usedin → MAESTRO
confidence 90% · A joint Widom insertion screen... then enforced three conditions simultaneously
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The coordination of multi-scale tasks is an effective strategy for computational materials discovery, yet the repeated application of diverse algorithms and tools renders it challenging. We report MAESTRO, a large language model (LLM) agent system capable of executing the entire screening pipeline for metal-organic frameworks (MOFs). It processes a large body of MOF literature, links relevant publications to their crystal structures, and curates the results into a computation-ready database, which is then screened through a strategy of progressively increasing computational cost. The promising candidates identified for separation under wet flue gas conditions all originate from unrelated studies. By connecting the heterogeneous stages of computational materials discovery, the LLM-based agents of MAESTRO can operate across application domains and uncover high-performance materials that conventional screening approaches would be unlikely to consider.
Tags
Links
- Source: https://arxiv.org/abs/2608.20434v1
- Canonical: https://arxiv.org/abs/2608.20434v1
Trouble viewing inline? Open PDF directly →
Full Text
40,382 characters extracted from source content.
Expand or collapse full text
1 An LLM agent for end-to-end computational materials discovery Chen Yuntong 1 , Huang Ju 3 , Liu Yu 4 , Zhao Dan 5 , Sun Mingqi 1 , Ju Chentian 6 , Liu Yanbing 4 , Huang Lijiang 1 * , Zhao Guobin 2 * 1 School of Mechanical Engineering, Northwestern Polytechnical University, Xi’an, 710072, Shaanxi, China. 2* Department of Chemistry, National University of Singapore, Singapore, 117543, Singapore. 3 Chemical Engineering and Applied Chemistry, University of Toronto, Toronto, M5S 3E5, ON, Canada. 4 Institute of Information Engineering, Chinese Academy of Sciences, Beijing, 100045, China. 5 Department of Radiology, Huazhong University of Science and Technology, Wuhan, 430022, Hubei, China. 6 Department of Complexity Science and Engineering, The University of Tokyo, Kashiwa, 277-8561, Chiba, Japan *Corresponding author(s). E-mail(s): haunglj@nwpu.edu.cn; sxmzhaogb@gmail.com; Contributing authors: chenyt123@mail.nwpu.edu.cn; huangju33@gmail.com; liuyu@iie.ac.cn; d202382054@hust.edu.cn; sun2024@mail.nwpu.edu.cn; Abstract The coordination of multi-scale tasks is an effective strategy for computational materials discovery, yet the repeated application of diverse algorithms and tools renders it challenging. We report MAESTRO, a large language model (LLM) agent system capable of executing the entire screening pipeline for metal-organic frameworks (MOFs). It processes a large body of MOF literature, links relevant publications to their crystal structures, and curates the results into a computation-ready database, which is then screened through a strategy of progressively increasing computational cost. The promising candidates identified for separation under wet flue-gas conditions all originate from unrelated studies. By connecting the heterogeneous stages of computational materials discovery, the LLM-based agents of MAESTRO can operate across application domains and uncover high-performance materials that conventional screening approaches would be unlikely to consider. 2 1. Introduction Artificial intelligence has markedly advanced the pace of materials discovery. 1-4 A complete discovery study, however, spans stages that differ in kind, from acquiring reliable starting structures to computing their properties and screening candidates through progressively expensive methods. 5-7 How to bridge these stages into a coherent research campaign remains a central challenge. Metal-organic frameworks (MOFs) are porous crystalline materials, assembled from metal nodes and organic linkers, 8 possess tunable pore geometry, chemistry environment that have driven applications in gas separation, 9 catalysis, 10 drug delivery. 11 To date, more than 100,000 experimentally determined structures have been deposited in crystallographic databases, 12 and in silico screening for specific applications has become a prominent research approach. 13-15 However, this requires computational chemistry researchers to master and coordinate a diverse set of skills and software tools. 16-18 Recently, large language models (LLMs) have been applied to individual stages of this discovery process within the MOF domain, 19-23 including extracting synthesis conditions from the literature, linking material names to crystallographic identifiers, predicting adsorption properties through conversational interfaces, and repairing or generating framework structures. In parallel, curated database efforts such as the CoRE MOF 2024 have demonstrated that structure cleaning, property assignment and application screening can be integrated into a single pipeline. 24 These two lines of progress, LLM-based reasoning for individual tasks and systematic multi-stage curation, have each proved effective, but they have developed largely independently of each other. 25,26 It remains untested whether an LLM-driven agent can coordinate such a multi-stage workflow, carrying a study from a scientific question through to validated candidate materials. We demonstrate that an LLM agent system (MAESTRO) for performing complete in silicon screening from scratch. The system processed the broad MOF literature, associated publications with crystal structures, curated the resulting collection into a database, and screened it through progressively expensive methods. We showed a case study for CO2 capture from wet flue gas and determined the top-performing candidates under humid conditions. 3 2. Results 2.1 Overview of MAESTRO MAESTRO employs a multi-agent architecture in which one orchestrator coordinates four specialist agents (Fig. 1a). Each specialist handles one phase of the study: literature filtering, structure collection and validation, property calculation, and multi-scale screening. The execution layer calls established computational tools to ensure reproducibility. The decision layer is driven by a large language model, with every decision recorded as a structured trace that can be audited and replayed. Each specialist operates through a plan-reflect-gate cycle (Fig. 1b). Before execution, the agent plans its strategy on the basis of the current data state. After execution, it diagnoses the quality of its own output and reports both results and diagnostics to the orchestrator, which then reasons about whether to proceed or request revision. Two examples illustrate this cycle. The structure agent, tasked with ordering five independent validation tools into an efficient pipeline, benchmarked each tool's throughput and rejection rate on a random sample and used the measurements to determine the execution sequence. The literature agent, after filtering the retrieved papers to a MOF-relevant subset, sampled accepted and rejected entries and diagnosed a precision of 77%. It identified studies in which MOFs served only as precursors for derived carbons as the dominant false-positive pattern and proposed exclusion rules. The orchestrator read this report, reasoned that the remaining false positives would be eliminated downstream when papers lacking crystal structures are naturally excluded, and issued a proceed decision. The literature agent designed query terms covering variants and synonyms of the MOF nomenclature and retrieved over 400,000 papers from OpenAlex. 27 It associated these papers with crystal structures from the Cambridge Structural Database (CSD) 28 and the Crystallography Open Database (COD), 29 achieving 92.1% bidirectional linkage. This literature layer does not restrict the screening entry point. Instead, it provides provenance tracking that links every structure to its original publication, enabling the cross-domain discovery reported below. The structure agent then cleaned by coremof-tools 24 and validated the crystal structures through a five-tool pipeline (Chen-Manz, 30 MOFChecker, 31 MOSAEC, 17 MOFClassifier, 16 SETC 32 ). After symmetry reduction and solvent removal, the agent produced over 190,000 candidate files covering both full-solvent-removed and atomic-solvent-removed variants. Each file was evaluated independently by all five validation tools, and only those approved by all five were retained. This conservative intersection strategy yielded 64,097 CR entries. Post-hoc analysis of structures rejected by only a single tool confirmed that the strategy favored database reliability over completeness. The property agent then computed 9 categories of structural and chemical descriptors for these structures, achieving near-complete coverage (Fig. 1c). 33-39 4 This end-to-end integration distinguishes MAESTRO from existing systems that each address only one or two stages of the discovery chain (Fig. 1d). To verify that the workflow supports continuous operation, we ran the complete pipeline on crystal structures deposited in the three months following the initial database construction without manual intervention. MAESTRO autonomously retrieved newly published papers, associated them with crystal structures, and processed the candidates through the same validation and featurization pipeline, incorporating ~300 new CIFs into the database. Fig. 1 Architecture, auditable decisions and data lineage of MAESTRO. a, MAESTRO architecture linking literature and structural resources to a large-language-model orchestrator, four specialized agents, a shared database, a typed message bus and a decision ledger. The orchestrator integrates context, routes tasks and applies quality gates, while the Literature, Structure, Property and Screening Agents interact with the shared state to produce a living MOF database and validated discoveries. b, Examples of agent decision traces. The Structure Agent uses a self-proposed pilot benchmark to plan the order of five structure checkers before full execution. The Literature Agent audits stratified samples of matched and rejected records after execution, diagnoses residual errors, tests a revised rule and reports the result to the orchestrator for a proceed-or-revise gate. The audit samples contain 100 matched and 100 rejected records and are not a population-level confusion matrix. c, Linked literature and structure data lineage. Paper records and CSD structures are connected bidirectionally through digital object identifiers for provenance but remain independent input populations. The structure track shows preprocessing, conversion to the P1 setting, generation of free solvent removal (FSR) and all solvent removal (ASR) variants, five-checker computation-readiness validation and property featurization. Counts refer to different entities, namely papers, structures, and structure-file variants, and therefore do not represent a single-denominator retention funnel. d, Capability landscape for 9 materials databases or artificial-intelligence systems across eight research-pipeline and three agentic-methodology criteria. 19,21,24,40-44 Light, intermediate and dark cells denote not demonstrated, partially demonstrated and demonstrated, respectively, in the cited report; operational definitions and primary-publication evidence are provided in the Supplementary Information. Source data are provided with this paper. 5 6 2.2 Virtual screening for carbon capture from wet flue gas Given the research objective of identifying MOFs suitable for CO2 capture from wet flue gas, the screening agent translated the objective into a seven-step workflow with specific quantitative criteria at each stage (Fig. 2a). The first three steps served as progressively expensive filters. A rule-based prefilter removed structures with pore diameters too small to accommodate CO2, those containing noble metals, and those with open metal sites (OMSs). OMS typically has a strong affinity for water and cannot be described by universal force fields. This step reduced the pool from 64,071 to 19,207. A joint Widom insertion screen at 313 K then enforced three conditions simultaneously: a water Henry coefficient below 5E-6 mmol/g/Pa to ensure thermodynamic hydrophobicity, 45 a CO2 Henry coefficient above 1E-6 mmol/g/Pa to ensure sufficient affinity for the target gas, and a CO2/N2 Henry selectivity exceeding 5. The candidate pool fell from 19,207 to 1,035 (Fig. 2b), substantially reducing the computational cost of the subsequent simulation step. Grand canonical Monte Carlo (GCMC) simulation 46,47 under a CO2/N2 mixture (15:85 molar ratio) at 313 K and 1 bar then identified 280 structures with a working capacity above 0.5 mol/kg. A geometric accessibility check was then applied to the 280 candidates. The prefilter in Step 1 had screened the largest cavity diameter (퐷 ! ), which measures whether a pore can host a CO2 molecule. It did not screen the pore-limiting diameter (퐷 " ), which measures whether the molecule can traverse the windows connecting adjacent cavities. GCMC places molecules by random insertion without simulating diffusion, and structures with large cavities but constricted windows therefore yield a thermodynamic upper bound rather than a kinetically accessible capacity. Applying a 퐷 " ≥ 2.4 Å threshold, chosen to exceed the diameter of the H2 molecule, removed 75 such structures and reduced the pool to 205 (Fig. 2c). Given that using the kinetic diameter of CO2 as a cutoff would exclude the rigid CALF-20 and other promising candidates, we have selected CALF-20 48 —a commercially available carbon capture material—as our benchmark, even though it was excluded based on its Henry’s constant for water in Step 3. From these, the screening agent selected 11 candidates spanning two complementary regions of the selectivity versus capacity design space (Fig. 2d). 7 Fig. 2 Agent-directed multi-fidelity screening of MOFs for wet-flue-gas CO2 capture. a, Screening Agent architecture translating a natural-language research objective into a strategy that queries the living property database and invokes molecular simulation, process modelling and machine-learned-potential tools before returning ranked candidates. b, Sequential Widom screen of 19,207 screening-ready structures. The criteria require a H2O Henry coefficient below 5E−6 mmol/g/Pa, a CO2 Henry coefficient above 1E−6 mmol/g/Pa and a Henry CO2/N2 selectivity above 5, leaving 1,035 structures. Card size is adjusted for legibility and exact counts are printed. c, Complete multi- fidelity funnel from five-checker CR entries through geometric and compositional pre-filtering, Widom screening, binary grand canonical Monte Carlo simulation, pore-accessibility analysis and detailed process simulation. Funnel width is logarithmically scaled. d, Working capacity versus selectivity for the 205 structures satisfying the static accessibility criterion. Grey circles show the complete accessible set, colored circles show the 12 MOFs. e, Maximum dry pressure-swing-adsorption CO2 purity at recovery of at least 90% for the 12 MOFs. Each value is selected from 3,600 valid process-simulation points for that material, and the dashed line marks the CALF-20 baseline. 8 9 2.3 Multi fidelity evaluation yields a converging top tier The 11 candidates and CALF-20 were then evaluated through four steps of increasing fidelity: heat capacity calculation, single-component isotherm fitting, vacuum swing adsorption (VSA) process simulation (Supplementary Information), and ternary competitive adsorption under humidity. The VSA simulation evaluated two independent objectives: a performance objective maximizing CO2 purity and recovery, and an economic objective minimizing energy consumption while maximizing productivity. Under the performance objective, ZAWZAA achieved 96.01% CO2 purity at ≥90% recovery, and 10 of the 11 candidates exceeded CALF-20 at 87.07% (Fig. 3a). Under the economic objective, YIWBEN dominated the global Pareto front, contributing all non-dominated solutions while no other material contributed a single point (Fig. 3b). It achieved the lowest energy consumption and the highest productivity among all 12 materials, leaving no trade-off space for competing candidates. The performance top three (ZAWZAA, YIWBEN, VICDOC) and the economic top three (YIWBEN, VICDOC, ZAWZAA) comprised the same set of materials in different internal order (Fig. 3c). This overlap indicates that the head of the ranking is robust to the choice of objective function. CALF-20 ranked 11th of 12 under the performance objective and 8th under the economic objective, confirming that even under dry conditions the majority of hydrophobic candidates already outperformed the commercial benchmark. The process simulation results also exposed a limitation of simpler screening proxies. An equilibrium-based ideal VSA calculation, which reflects the thermodynamic upper bound of single-stage separation, diverged substantially from the full process ranking: YIWBEN rose from 10th by the ideal metric to 2nd in the full simulation, while BELCUT fell from 1st to 6th (Fig. 3d). Had the ideal metric been used to prune candidates before process simulation, the economic champion would have been discarded. The preceding evaluation was performed under dry conditions and therefore did not account for competitive water adsorption. To test whether the hydrophobic screening criterion translates into actual humidity tolerance, we evaluated all 12 materials through GCMC simulation (CO2/N2/H2O) across 11 relative humidity (RH) levels from 0 to 1.0 at 313 K and 1 bar. All 11 hydrophobic candidates retained between 93.7% and 98.8% of their dry-state CO2 capacity at full water saturation (RH = 1.0), with water uptake below 0.047 mol kg⁻¹ (Fig. 4a). CALF-20 absorbed 9.79 mol kg⁻¹ of water under the same conditions, and its CO2 capacity collapsed to 6.5% of the dry-state value (Fig. 4b). The collapse was not gradual: CALF-20 maintained nearly 90% of its CO2 capacity up to RH ≈ 0.7, but beyond this point water uptake surged and CO2 retention dropped sharply, consistent with capillary 10 condensation in its hydrophilic pore network (Fig. 4c). This result does not imply that CALF-20 is unsuitable for CO2 capture, but rather that its operating window is bounded at approximately RH 0.7. Three independent lines of evidence thus converged on the same conclusion. Widom insertion identified low water affinity at zero loading. Dry-state process simulation confirmed superior separation performance. Ternary competitive adsorption demonstrated humidity tolerance under saturation. Each method is physically distinct, yet all pointed consistently to the same set of candidates outperforming the commercial benchmark. 11 Fig. 3 Process-level and economic evaluation of shortlisted MOFs. a, Material-specific vacuum-swing- adsorption Pareto frontiers in CO2 recovery–purity space. The complete dataset contains 3,600 valid process- simulation points for each of 12 materials (43,200 points in total); the vertical line marks the 90% recovery requirement used for performance ranking. b, Economic Pareto frontiers relating energy consumption to productivity. Points are the final 60 feasible optimization solutions for each material (720 points in total); coloured lines show material-level non-dominated fronts, and the overlaid frontier denotes non-dominated solutions after pooling all materials. Endpoint annotations identify the minimum-energy and maximum-productivity solutions on the global frontier. c, Paired rank flow for the same 12 materials, connecting the process-performance rank based on the highest purity at recovery of at least 90% to the economic rank based on the final compromise solution. Ribbon width provides visual emphasis only and does not encode a quantitative variable; the shaded regions mark the top 3 positions. d, Consistency between ideal vacuum-swing-adsorption ranks and full pressure-swing-adsorption ranks for the 12 materials. The dashed diagonal denotes rank preservation; arrows terminate at the full-process rank and shaded bands mark the respective top 3 regions. Source data are provided with this paper. 12 Fig. 4 Competitive adsorption responses across relative humidity. a, H2O uptake versus CO2 uptake for 11 discovered candidates and the external CALF-20 baseline over relative humidity (RH) values from 0.1 to 1.0 (120 material–condition observations). Colors identify materials and marker area increases with RH. Horizontal and vertical error bars are the RASPA block-averaging errors reported for the H2O and CO2 uptake estimates, respectively. b, CO2 retention at RH = 1 relative to the dry-state uptake for the same 12 materials. Retention is calculated as CO2 uptake at RH = 1 divided by CO2 uptake at RH = 0, multiplied by 100%. The dashed line marks the 90% criterion, and the horizontal axis is broken between 20% and 90% to show CALF-20 and the candidates without compressing their differences. Bars show ratios of the archived central estimates; the corresponding dry and wet uptake errors are provided in the Source Data and are not propagated under an unverified independence assumption. c, H2O uptake across RH = 0.1–1.0 for all 12 materials (120 material–condition observations). Vertical error bars are the RASPA block-averaging errors reported for individual simulations. In a and c, the 12 dry-state observations have zero H2O uptake and are omitted only because the H2O axes are logarithmic; no wet-state observation is aggregated, imputed or excluded. Source data are provided with this paper. 13 14 2.4 Robust candidates emerge beyond application-driven screening Provenance tracking through the literature layer revealed the research origins of all structures that entered the screening pipeline. Among the 49,522 unique framework refcodes in the validated collection, the largest source domains were luminescence and sensing, structure reports and coordination chemistry, each contributing roughly one-fifth of the pool. CO2 capture and separation accounted for only 1.0% (Fig. 5a). All 11 candidates originated from studies unrelated to CO2 capture (Fig. 5b). ZAWZAA, the performance champion, was first reported in a study of luminescent Schiff-base coordination polymers. YIWBEN, the economic champion, came from a magnetic relaxation study of a Co-MOF. VICDOC, ranked third under both objectives, appeared in work on hexane isomer separation in Fe-MOF triangular channels. Four of the 11 candidates (LAXMOP, KAXDAR, SOYZOY, YIWBEN) traced to magnetism research, suggesting that ligand choices in magnetic MOF design may incidentally favor hydrophobic pore environments. The remaining candidates spanned photocatalysis, coordination chemistry and crystal structure characterization (Supplementary Information). An application-oriented literature search targeting CO2 capture keywords would not have surfaced any of them. A validation matrix summarizes the convergence of evidence across the three independent methods (Fig. 5c). Ten of the 11 candidates passed all three criteria: Henry coefficient hydrophobicity, dry-state purity exceeding the CALF-20 baseline, and wet-state CO2 retention above 90%. BEFJOO passed two of three. Its dry purity of 79.50% fell below that of CALF-20 (87.07%), yet it retained 95.86% of its CO2 capacity under full humidity, confirming that hydrophobicity and process-level separation performance are independent dimensions. CALF-20 itself passed none of the three criteria. To assess the rigid-framework assumption underlying all GCMC simulations, we performed machine- learning potential molecular dynamics (MLP-MD) simulations using the UMA-S-1.1 interatomic potential for all 12 structures under four guest conditions: empty framework, CO2, N2 and H2O. Nine structures satisfied the rigidity criterion, with dynamic pore diameter variation below 10% of the static value across all conditions (Fig. 5d). Among the three structures showing larger deviations, each case was physically interpretable. YIWBEN exhibited framework contraction when empty but recovered under CO2 loading, with a mean pore diameter 34% larger than in the empty state, consistent with guest-induced stabilization. SOYZOY showed reduced dynamic accessibility under multiple conditions but retained sufficient pore diameter for CO2 passage. KAXDAR experienced an 18% contraction under CO2 specifically, warranting experimental validation of its transport properties. In all three cases the pore diameter remained above the accessibility threshold under the guest conditions relevant to the screening results. 15 These results demonstrate that an LLM-driven agent can bridge the heterogeneous stages of computational materials discovery. By operating across application domains rather than within a single literature silo, the system surfaced high-performance candidates that conventional screening campaigns are unlikely to consider. 16 Fig. 5 Research-domain provenance and multi-fidelity validation of discovered MOFs. a, Primary research domains of 49,522 unique framework refcodes, assigned from source-paper titles and abstracts. Free-solvent- removed and all-solvent-removed variants of the same refcode are counted once. Domains below 1% are grouped as other applications; CO2 capture and separation remain separate as the scientific contrast. b, Source domains of the 11 discovered candidates. Each equal-area tile represents one candidate and uses the domain color in a; the open diamond denotes the external CALF-20 baseline. c, Multi-fidelity validation matrix for the candidates and CALF-20. Columns report the H2O Henry coefficient, maximum dry pressure-swing-adsorption purity at recovery of at least 90%, and CO2 retention at RH = 1. Pass criteria are below 5 × 10−6 mmol g−1 Pa−1, above the CALF-20 dry-purity baseline, and at least 90%, respectively. Colour scales are normalized within columns, with darker cells indicating better performance; the rightmost column gives the number of passed levels. d, Dynamic pore-accessibility matrix from 12,046 archived machine-learned-potential molecular-dynamics frames. Bubble colour encodes mean dynamic free- sphere diameter (Df) change relative to the same material's empty trajectory and is clipped from −20% to +15%. Bubble area is proportional to the percentage of frames with Df ≥ 3.05 Å. This threshold is the O-site Lennard– Jones sigma of the TraPPE CO2 model and indicates force-field-informed geometric clearance, not diffusion. There are 251 frames for each of 46 material–condition groups and 250 frames for BEFJOO–CO2 and CALF-20–N2. Source data are provided with this paper. 17 18 3. Discussion TMAESTRO demonstrates that an LLM agent system can execute a complete computational materials discovery study, coordinating heterogeneous stages from literature processing through validated candidate selection. Rather than automating a single task or providing a conversational interface to existing tools, the system carried a research campaign from a scientific question to candidate materials validated through three independent methods. The resulting candidates and the methodological observations reported above can be evaluated on their own scientific merits, independent of the agent infrastructure that produced them. The hydrophobic screening strategy yielded 11 candidates that retained over 93% of their dry-state CO2 capacity at full water saturation, while CALF-20 retained only 6.5%. This contrast reflects two fundamentally different design strategies for humid-condition CO2 capture. CALF-20 pairs hydrophilic pores with exceptional water stability, maintaining performance up to approximately RH 0.7 before capillary condensation degrades uptake. The hydrophobic candidates instead present pore environments that are thermodynamically unfavorable for water adsorption, avoiding competitive binding altogether. These strategies are complementary rather than competing: the choice between them depends on the humidity conditions of the target application. Experimental synthesis and characterization of the top candidates will be necessary to confirm the computationally predicted performance. Several limitations should be noted. All adsorption simulations assumed rigid frameworks and used classical force fields. MLP-MD validation confirmed that 9 of 12 structures maintained their pore geometry under guest loading, but YIWBEN showed guest-induced stabilization that deviates from the rigid approximation, and KAXDAR exhibited CO2-specific contraction. Both cases warrant experimental transport measurements. In addition, the agent's decisions were generated by a single LLM. Different models may produce different strategy choices at the planning and reflection stages, and the sensitivity of the final outcomes to model selection has not been systematically evaluated. The structured decision traces recorded throughout the pipeline provide a basis for such evaluation in future work. Finally, the reflect mechanism relies on sampling-based audits rather than exhaustive verification, and residual errors in the curated database cannot be fully excluded. 4. Methods 4.1 Agent architecture and decision tracing MAESTRO consists of one orchestrator agent and four specialist agents responsible for literature retrieval, structure curation, property computation and multi-fidelity screening, respectively. The agents 19 communicate through a typed message bus rather than direct calls. The orchestrator dispatches task requests to one specialist at a time, receives the execution result together with a self-diagnostic report and decides whether to proceed to the next phase or request revision. Each specialist operates through a plan-reflect-gate cycle. In the plan stage, the agent examines the current data state and formulates an execution strategy. In the execute stage, the agent calls deterministic computational tools without intervention from the language model, ensuring that all numerical outputs are reproducible. In the reflect stage, the agent diagnoses the quality of its own output by sampling, auditing or cross-checking results. The orchestrator reads the diagnostic report, reasons about whether the phase has met its quality criteria and issues a proceed or halt decision at the gate. Every decision is recorded as a structured JSON trace containing three fields: the context presented to the model, the reasoning chain and the final decision. These traces are persisted to disk and enable full audit and replay of the decision history. Representative traces and the complete JSON schema are provided in Supplementary Information. All agent decisions were generated by Claude Opus 4.6 (Anthropic) accessed through the official API. Inference parameters and all computational tool versions are reported in Supplementary Information. 4.2 Autonomous literature processing and structure curation The literature agent autonomously designed query terms, applied regex-based filtering rules to title and abstract fields and verified filter quality by sampling from both accepted and rejected pools. It diagnosed precision and recall using the language model and proposed rule revisions, which were tested on a held-out sample before adoption. The orchestrator reviewed the diagnostic report before issuing a proceed decision. The complete keyword list and filtering rules are provided in Supplementary Information. All acquired structures were converted to P1 symmetry and processed through two parallel solvent-removal channels. Five independent validation tools were then applied in sequence. To determine an efficient execution order, the structure agent benchmarked each tool on a random sample, measuring throughput and rejection rate, and used the measurements to arrange the pipeline. A structure was labelled CR only if all five tools approved it. The property agent then computed nine categories of structural and chemical descriptors for each CR structure. All tool names, versions and module configurations are provided in Supplementary Information. 4.3 Agentic multi-fidelity screening Given a natural-language research objective, the screening agent translated it into a seven-step workflow with quantitative criteria at each stage. The agent autonomously determined filter thresholds, simulation 20 conditions and candidate selection logic. Steps were ordered by increasing computational cost so that inexpensive filters removed the bulk of candidates before expensive simulations were applied. After the initial screening round, the reflect mechanism diagnosed a geometric accessibility issue in the selected candidates. The agent proposed an additional pore-window check, which was verified and adopted, leading to reconstruction of the candidate pool. This self-correction is described in Section 2.2. All simulation tools, force fields, thermodynamic conditions, screening thresholds and ranking definitions for the subsequent evaluation stages are reported in Supplementary Information. 4.4 Continuous operation To verify that the system supports continuous operation, the complete pipeline was re-executed on crystal structures deposited during the three months following the initial database construction. No manual intervention was applied. The orchestrator autonomously coordinated the literature, structure and property agents to retrieve newly published papers, associate them with newly deposited structures and process candidates through the same validation and featurization pipeline. A screen recording of the system executing an incremental update cycle is provided as Supplementary Video. 21 References 1 Stein, H. S. & Gregoire, J. M. Progress and prospects for accelerating materials science with automated and autonomous workflows. Chemical science 10, 9640–9649 (2019). 2 Batra, R., Song, L. & Ramprasad, R. Emerging materials intelligence ecosystems propelled by machine learning. Nature Reviews Materials 6, 655–678 (2021). 3 Pyzer-Knapp, E. O. et al. Accelerating materials discovery using artificial intelligence, high performance computing and robotics. npj Computational Materials 8, 84 (2022). 4 Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023). 5 Charalambous, C. et al. A holistic platform for accelerating sorbent-based carbon capture. Nature 632, 89–94 (2024). 6 Xiang, Z., Cao, D., Lan, J., Wang, W. & Broom, D. P. Multiscale simulation and modelling of adsorptive processes for energy gas storage and carbon dioxide capture in porous coordination frameworks. Energy & Environmental Science 3, 1469–1487 (2010). 7 Deng, Z. et al. Towards autonomous high-throughput multiscale modelling of battery interfaces. Energy & Environmental Science 15, 579–594 (2022). 8 Furukawa, H., Cordova, K. E., O’keeffe, M. & Yaghi, O. M. The chemistry and applications of metal- organic frameworks. Science 341, 1230444 (2013). 9 Su, W. et al. Challenges and recent advances in MOF-based gas separation membranes. Chemical Communications 60, 7124–7135 (2024). 10 Bavykina, A. et al. Metal–organic frameworks in heterogeneous catalysis: recent progress, new trends, and future perspectives. Chemical reviews 120, 8468–8535 (2020). 11 Wu, M. X. & Yang, Y. W. Metal–organic framework (MOF)‐based drug/cargo delivery and cancer therapy. Advanced Materials 29, 1606134 (2017). 12 Moghadam, P. Z. et al. Targeted classification of metal–organic frameworks in the Cambridge structural database (CSD). Chemical science 11, 8373–8387 (2020). 13 Boyd, P. G. et al. Data-driven design of metal–organic frameworks for wet flue gas CO2 capture. Nature 576, 253–256 (2019). 14 Jablonka, K. M., Ongari, D., Moosavi, S. M. & Smit, B. Big-data science in porous materials: materials genomics and machine learning. Chemical reviews 120, 8066 (2020). 15 Wilmer, C. E. et al. Large-scale screening of hypothetical metal–organic frameworks. Nature chemistry 4, 83–89 (2012). 16 Zhao, G., Zhao, P. & Chung, Y. G. Mofclassifier: a machine learning approach for validating computation-ready metal–organic frameworks. Journal of the American Chemical Society 147, 33343– 33349 (2025). 17 White, A. J., Gibaldi, M., Burner, J., Mayo, R. A. & Woo, T. K. High structural error rates in “computation-ready” MOF databases discovered by checking metal oxidation states. Journal of the American Chemical Society 147, 17579–17583 (2025). 18 Gibaldi, M. et al. MOSAEC-DB: a comprehensive database of experimental metal–organic frameworks with verified chemical accuracy suitable for molecular simulations. Chemical Science 16, 4085–4100 (2025). 19 Pruyn, T. M. et al. Mof-chemunity: Literature-informed large language models for metal–organic framework research. Journal of the American Chemical Society 147, 43474–43486 (2025). 20 Ozcan, A. et al. Artificial intelligence paradigms for next-generation metal–organic framework research. Journal of the American Chemical Society 147, 23367 (2025). 21 Kang, Y. & Kim, J. ChatMOF: an artificial intelligence system for predicting and generating metal- organic frameworks using large language models. Nature communications 15, 4705 (2024). 22 Zheng, Z. et al. Shaping the water-harvesting behavior of metal–organic frameworks aided by fine- tuned GPT models. Journal of the American Chemical Society 145, 28284–28295 (2023). 22 23 Jablonka, K. M., Schwaller, P., Ortega-Guerrero, A. & Smit, B. Leveraging large language models for predictive chemistry. Nature Machine Intelligence 6, 161–169 (2024). 24 Zhao, G. et al. CoRE MOF DB: A curated experimental metal-organic framework database with machine-learned properties for integrated material-process screening. Matter 8 (2025). 25 Chen, Z. et al. An agentic artificially intelligent X-ray scientist. Nature Machine Intelligence, 1–12 (2026). 26 Kim, Y. et al. Capable language models can outgrow the benefits of collaboration. Nature Machine Intelligence 8, 1157–1172 (2026). 27 Priem, J., Piwowar, H. & Orr, R. OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv preprint arXiv:2205.01833 (2022). 28 Groom, C. R., Bruno, I. J., Lightfoot, M. P. & Ward, S. C. The Cambridge structural database. Structural Science 72, 171–179 (2016). 29 Gražulis, S. et al. Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration. Nucleic acids research 40, D420–D427 (2012). 30 Chen, T. & Manz, T. A. Identifying misbonded atoms in the 2019 CoRE metal–organic framework database. RSC advances 10, 26944–26951 (2020). 31 Jin, X., Jablonka, K. M., Moubarak, E., Li, Y. & Smit, B. MOFChecker: a package for validating and correcting metal–organic framework (MOF) structures. Digital Discovery 4, 1560–1569 (2025). 32 Gibaldi, M. et al. Generalizable classification of crystal structure error types using graph attention networks. Journal of Materials Chemistry A 13, 32255–32270 (2025). 33 Ong, S. P. et al. Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis. Computational Materials Science 68, 314–319 (2013). 34 Hjorth Larsen, A. et al. The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter 29, 273002 (2017). 35 Zoubritzky, L. & Coudert, F.-X. CrystalNets. jl: Identification of crystal topologies. SciPost Chemistry 1, 005 (2022). 36 Willems, T. F., Rycroft, C. H., Kazi, M., Meza, J. C. & Haranczyk, M. Algorithms and tools for high- throughput geometry-based analysis of crystalline porous materials. Microporous and Mesoporous Materials 149, 134–141 (2012). 37 Bucior, B. J. et al. Identification schemes for metal–organic frameworks to enable rapid search and cheminformatics analysis. Crystal Growth & Design 19, 6682–6697 (2019). 38 Chung, Y. G. et al. Advances, updates, and analytics for the computation-ready, experimental metal– organic framework database: CoRE MOF 2019. Journal of Chemical & Engineering Data 64, 5985–5998 (2019). 39 Jablonka, K. M., Ongari, D., Moosavi, S. M. & Smit, B. Using collective knowledge to assign oxidation states of metal cations in metal–organic frameworks. Nature Chemistry 13, 771–777 (2021). 40 Inizan, T. J. et al. System of agentic AI for the discovery of metal-organic frameworks. arXiv preprint arXiv:2504.14110 (2025). 41 Kang, Y. et al. Harnessing large language models to collect and analyze metal–organic framework property data set. Journal of the American Chemical Society 147, 3943–3958 (2025). 42 Wang, K. et al. Synthesis of Highly Crystalline Covalent Organic Frameworks Using Large Language Models. Journal of the American Chemical Society 148, 10094–10104 (2026). 43 Chaudhari, A., Ock, J. & Barati Farimani, A. Modular large language model agents for multi-task computational materials science. Communications Materials 7, 131 (2026). 44 Ding, K. et al. SciToolAgent: a knowledge-graph-driven scientific agent for multitool integration. Nature Computational Science 5, 962–972 (2025). 45 Moghadam, P. Z., Fairen-Jimenez, D. & Snurr, R. Q. Efficient identification of hydrophobic MOFs: application in the capture of toxic industrial chemicals. Journal of Materials Chemistry A 4, 529–536 (2016). 46 Ran, Y. A. et al. RASPA3: A Monte Carlo code for computing adsorption and diffusion in nanoporous materials and thermodynamics properties of fluids. The Journal of Chemical Physics 161 (2024). 23 47 Dubbeldam, D., Torres-Knoop, A. & Walton, K. S. On the inner workings of Monte Carlo codes. Molecular Simulation 39, 1253–1292 (2013). 48 Lin, J.-B. et al. A scalable metal-organic framework as a durable physisorbent for carbon dioxide capture. Science 374, 1464–1469 (2021).