Source-linked AI summary
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
TL;DR
Computational materials discovery is difficult because it requires coordinating heterogeneous stages and diverse tools, while prior LLM applications largely addressed individual tasks. MAESTRO connects literature processing, crystal-structure linking, database curation, and progressively expensive screening, identifying wet-flue-gas CO2-capture candidates with strong retained capacity.
Problem
Computational materials discovery requires coordinating diverse algorithms, tools, and workflow stages, while prior LLM applications and curation efforts largely developed separately.
Method
MAESTRO processes MOF literature, links publications to crystal structures, curates a computation-ready database, and screens it through progressively expensive methods.
Results
Over 93% of dry-state CO2 capacity was retained at full water saturation by 11 hydrophobic-screening candidates, compared with 6.5% for CALF-20.
Takeaways & Limitations
MAESTRO carried a research campaign from a scientific question to candidate materials validated through three stages and surfaced candidates from unrelated studies.
Takeaways & Limitations
Adsorption simulations assumed rigid frameworks, and the agent's decisions were generated by a single LLM.
Abstract
from arXiv · showhide
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.
1. Introduction
Computational materials discovery requires coordinating diverse stages, algorithms, and software, but existing LLM and curation efforts have largely addressed separate parts of the workflow. MAESTRO connects literature processing, structure curation, database construction, and progressively expensive screening for MOFs, identifying top candidates for wet-flue-gas CO2 capture.
- 1. Introduction: MOF discovery spans acquiring reliable starting structures, computing material properties, and screening candidates, making coordination across stages a central challenge.
- 1. Introduction: LLMs have been applied to individual MOF tasks, including extracting synthesis conditions, linking material names to crystal structures, and predicting adsorption properties.
- 1. Introduction: LLM-based reasoning and systematic multi-stage curation have each proved effective, but they have developed largely independently of each other.
- 1. Introduction: MAESTRO executes complete in silico MOF screening by processing literature, associating publications with crystal structures, curating a database, and applying progressively expensive methods.
- 1. Introduction: The system identifies top-performing candidates for CO2 capture from wet flue gas through this integrated screening pipeline.
2. Results
MAESTRO integrates literature retrieval, structure curation, validation, property computation, and progressively costly screening through auditable LLM-agent coordination. Applied to wet-flue-gas CO2 capture, it identified robust candidates largely absent from application-driven searches, including materials from unrelated research domains.
- MAESTRO architecture: MAESTRO uses an orchestrator and specialist agents to coordinate literature filtering, structure collection, validation, property calculation, and multi-scale screening.Each decision is recorded as a structured trace that can be audited and replayed.
- Literature and structure curation: Over 400,000 papers were retrieved and linked to crystal structures from the CSD and COD with 92.1% association performance.The resulting literature layer tracks provenance from each structure to its original publication.
- Literature and structure curation: ~300 new CIFs were incorporated into the database through autonomous retrieval, structure association, validation, and featurization.The workflow also produced over 190,000 candidate files and retained 64,097 after five-checker computation-readiness validation.
- Virtual screening: The progressively costly screening reduced the pool from 64,071 to 19,207 and then to 1,035 structures before identifying 11 candidates spanning complementary selectivity-capacity regions.A kinetic-diameter cutoff was avoided because it would exclude rigid CALF-20 and other promising candidates.
- Process and economic evaluation: Under the performance objective, ZAWZAA achieved 96.01% CO2 purity at ≥90% recovery, while 10 of 11 candidates exceeded CALF-20 at 87.07%.Under the economic objective, YIWBEN contributed all non-dominated solutions and achieved the lowest energy consumption and highest productivity.
- Humidity tolerance: At RH = 1.0, all 11 hydrophobic candidates retained 93.7%-98.8% of dry-state CO2 capacity, whereas CALF-20 retained 6.5%.The candidates had water uptake below 0.047 mol kg⁻¹, while CALF-20 absorbed 9.79 mol kg⁻¹ under the same conditions.
- Cross-domain discovery: All 11 candidates came from studies unrelated to CO2 capture, and four traced to magnetism research, demonstrating cross-domain provenance.ZAWZAA came from luminescent coordination-polymer research, YIWBEN from magnetic relaxation, and VICDOC from hexane-isomer separation.
3. Discussion
MAESTRO connects heterogeneous stages of computational materials discovery into an end-to-end campaign, producing candidates with strong humid-condition CO2 performance. Its results also expose scope boundaries involving experimental validation, framework flexibility, model dependence, and residual curation errors.
- MAESTRO coordinates literature processing, structure curation, property computation, and multi-fidelity screening from a scientific question to validated candidate materials.
- 11 candidates retained over 93% of their dry-state CO2 capacity at full water saturation, whereas CALF-20 retained only 6.5%.
- The hydrophobic and water-stable strategies are complementary, with the preferred design depending on target humidity conditions.
- Experimental synthesis and characterization remain necessary to confirm the computationally predicted performance.
- The simulations assumed rigid frameworks and classical force fields, although MLP-MD found that 9 of 12 structures maintained pore geometry under guest loading.
- Model-selection sensitivity was not systematically evaluated, and sampling-based audits cannot fully exclude residual errors in the curated database.
4. Methods
MAESTRO uses an orchestrator and specialist agents to coordinate literature retrieval, structure curation, property computation, and progressively costly screening. Agents plan, execute deterministic tools, reflect through audits, and gate progression while recording decisions for replay.
- MAESTRO consists of one orchestrator and four specialist agents covering literature retrieval, structure curation, property computation, and multi-fidelity screening.
- The orchestrator dispatches typed task requests, receives execution results and diagnostics, and decides whether to proceed or request revision.
- Each specialist follows a plan-reflect-gate cycle, calling deterministic computational tools during execution without language-model intervention.
- Literature agents generate query terms, apply regex filters, sample accepted and rejected records, diagnose precision and recall, and test revised rules before adoption.
- Structures are converted to P1 symmetry, processed through solvent-removal channels, benchmarked across validation tools, and retained only after approval by all five tools.
- The screening agent translates a natural-language objective into a seven-step workflow with autonomous thresholds, conditions, selection logic, and increasing computational cost.
- A reflect mechanism identified a geometric accessibility issue, added a verified pore-window check, and reconstructed the candidate pool.
- The pipeline was re-executed on structures deposited during the following three months without manual intervention, processing updates through the same validation and featurization workflow.