PaperlayerView exact revision on arXiv

Source-linked AI summary

TopoMAS: Large Language Model Driven Topological Materials Multiagent System

Baohua Zhang, Xin Li, Huangchao Xu, Zhong Jin, Quansheng Wu, Ce Li

arXiv:2507.04053v1cond-mat.mtrl-scics.AI

TL;DR

Topological-material discovery requires complex coordination across analysis, computation, and validation, while current AI systems have limited coverage of that workflow. TopoMAS organizes the process with hierarchical agents and knowledge-graph updates; the authors report SrSbO3 discovery and benchmark results for Qwen2.5-72B. The conclusion also identifies broader computational coverage and improved fault tolerance as future work.

  • Problem

    Topological-material characterization requires complex coordination, and existing AI systems do not adequately address its analyses or cover the comprehensive discovery workflow.

    Source: §1 · pp. 1, 3

  • Method

    TopoMAS is a hierarchical multi-agent framework that coordinates retrieval, reasoning, computation, and knowledge-graph updates in a closed-loop workflow.

    Source: §1 · p. 3, §2.1 · p. 3

  • Results

    TopoMAS guided discovery of SrSbO3, and Qwen2.5-72B achieved over 94% accuracy on structured tasks and a 7.46/10 composite score on open-ended tasks.

    Source: §1 · p. 3, §4 · p. 12

  • Takeaways & Limitations

    The authors present TopoMAS as supporting on-demand design of topological materials through an integrated workflow and report strong model compatibility across their benchmarks.

    Source: §4 · p. 12

    Future work will extend the framework beyond topological properties and improve fault tolerance; TopoOQ responses also take 91–207 seconds on average.

    Source: §3.4 · p. 8, §4 · p. 12

Abstract

from arXiv · show

Topological materials occupy a frontier in condensed-matter physics thanks to their remarkable electronic and quantum properties, yet their cross-scale design remains bottlenecked by inefficient discovery workflows. Here, we introduce TopoMAS (Topological materials Multi-Agent System), an interactive human-AI framework that seamlessly orchestrates the entire materials-discovery pipeline: from user-defined queries and multi-source data retrieval, through theoretical inference and crystal-structure generation, to first-principles validation. Crucially, TopoMAS closes the loop by autonomously integrating computational outcomes into a dynamic knowledge graph, enabling continuous knowledge refinement. In collaboration with human experts, it has already guided the identification of novel topological phases SrSbO3, confirmed by first-principles calculations. Comprehensive benchmarks demonstrate robust adaptability across base Large Language Model, with the lightweight Qwen2.5-72B model achieving 94.55% accuracy while consuming only 74.3-78.4% of tokens required by Qwen3-235B and 83.0% of DeepSeek-V3's usage--delivering responses twice as fast as Qwen3-235B. This efficiency establishes TopoMAS as an accelerator for computation-driven discovery pipelines. By harmonizing rational agent orchestration with a self-evolving knowledge graph, our framework not only delivers immediate advances in topological materials but also establishes a transferable, extensible paradigm for materials-science domain.

1 Introduction

Topological-material discovery combines demanding theory, computation, and validation, while existing AI systems leave gaps in end-to-end workflow coverage and cost efficiency. TopoMAS proposes a hierarchical multi-agent workflow and reports both benchmark performance from smaller models and discovery of SrSbO3.

1 Introduction

  • Exhaustive screening and large-scale first-principles calculations make topological-material discovery cross-disciplinary, while information silos can cause promising candidates to be overlooked.The traditional workflow proceeds from theoretical prediction to computational validation.Source: §1 · p. 1
  • Current AI systems have limited coverage of comprehensive materials discovery, and reliance on large-scale language models can make computational costs prohibitively high.The paper identifies reducing resource consumption without compromising performance as a system-design challenge.Source: §1 · p. 2
  • Topological-material characterization requires coordinated band-structure, symmetry, invariant, and property analyses that current AI systems have not adequately addressed.The paper aims to explore whether coordinated smaller models can handle these tasks cost-effectively.Source: §1 · pp. 2–3
  • TopoMAS uses hierarchical task decomposition and multi-level agent collaboration to evaluate whether smaller models can match or surpass larger LLMs at lower computational cost.The authors report extensive evaluations across benchmark datasets with different complexities.Source: §1 · p. 3
  • The system connects natural-language retrieval and reasoning, DFT calculations, results analysis, and knowledge-graph updates, and its workflow guided discovery of SrSbO3.The paper presents this as an end-to-end workflow assisting human scientists.Source: §1 · p. 3

2 Methodology

TopoMAS combines hierarchical agents and multiple data resources in a closed-loop system for topological-material research. Specialized agents retrieve and reason over materials information, generate structures, and coordinate property computations whose validated results can update the knowledge graph.

2.1 Overview

  • TopoMAS organizes research as a perception–reasoning–execution–feedback loop, with a Core Agent coordinating five specialized agents and task-level tools.Its resources span material databases, literature, TopoKG, first-principles software, and crystal-structure generation.Source: §2.1 · p. 3

2.2 Data Source

  • Five complementary resources provide materials data, literature, TopoKG knowledge, first-principles software, and crystal-structure generation capabilities.TopoKG contains over 28,000 materials and 37 node types.Source: §2.1 · p. 3, §2.2 · p. 3

2.3 Material Retrieval & Reasoning Subsystem

  • Literature search combines semantic matching, community detection, and centrality ranking, while KG queries translate natural-language requests into Cypher through template-guided optimization.The MG Agent also uses Conv-CDVAE and CrystalFormer for conditional crystal-structure generation.Source: §2.3 · pp. 4–5
  • A system-level ReAct workflow decomposes user queries, dispatches subtasks to specialized agents, and iteratively refines results using feedback and interaction history.The MP Agent adds a nested ReAct layer for API selection and targeted filtering that limits upstream information overload.Source: §2.3 · p. 5

2.4 Material Property Computation subsystem

  • The computation subsystem plans workflows, selects agents according to task dependencies, executes tools, and reviews outputs for file integrity, convergence, and accuracy.Detected issues trigger targeted recommendations and another planning-selection-action cycle.Source: §2.4 · p. 5
  • After tasks are validated, the RP Agent extracts structured results and incrementally updates TopoKG for continued knowledge evolution.The system also uses checkpoint and restart capabilities for extended DFT calculations.Source: §2.4 · p. 6

2.5 Software and Hardware

  • The system uses SymTopo to prepare topological-material inputs and analyze properties, VASP for DFT, and the Era supercomputer for computation..Source: §2.5 · p. 6

2.6 Web interface and visualization

  • The web interface provides material discovery and property-computation modules, while validated results stored in TopoKG can be returned directly for later queries instead of repeating calculations.The discovery dashboard supports natural-language queries and structured searches of computational task archives.Source: §2.6 · p. 6

3.1 Experiment details

The experiments use three domain-specific datasets and evaluate models with task-appropriate measures. LLM4Mat-Bench covers property prediction, while TopoQA and TopoOQ assess question answering and open-ended material inquiries.

3.1.1 Datasets

  • The test set contains 8,579 cleaned Materials Project structures and covers eight regression properties and two classification properties, alongside 110 TopoQA questions and 100 open-ended TopoOQ questions.The LLM4Mat-Bench source includes composition, CIF, and description representations.Source: §3.1.1 · pp. 6–7

3.1.2 Models and Evaluation Metrics

  • Across TopoMAS datasets, LLM4Mat-Bench used Qwen2.5-72B only (weighted MAD:MAE for eight regression tasks and AUC for two classification tasks), while TopoQA and TopoOQ evaluated all three models.TopoOQ combined a 0–10 composite quality score covering accuracy, completeness, and factual consistency with Average Time, Median Tokens, and Total Tokens efficiency metrics.Source: §3.1.2 · p. 7

3.2 Benchmark on LLM4Mat-Bench dataset

On LLM4Mat-Bench, TopoMAS based on Qwen2.5-72B outperforms the reported MatBERT-109M baseline in both regression and classification. The authors attribute the gains to specialized MP Agent tools and dynamic tool selection.

3.2 Benchmark on LLM4Mat-Bench dataset

  • Six specialized MP Agent tools and dynamic tool selection are credited with translating natural-language queries into optimized API workflows.The tools cover material summary, thermodynamics, elasticity, dielectric, electronic, and magnetic properties.Source: §3.2 · p. 8
  • Regression performance improves 168.4% to a weighted-average MAD:MAE of 14.421, versus MatBERT-109M’s 5.37.The scores are reported for the MP subset; higher MAD:MAE is better.Source: Table 1 · p. 8, §3.2 · p. 8
  • Classification improves 23.4% to a weighted-average AUC of 0.8914, versus MatBERT-109M’s 0.722.The authors also report AUCs of 0.92 for is_stable and 0.86 for is_gap_direct.Source: Table 1 · p. 8, §3.2 · p. 8

3.3 Benchmark on TopoQA dataset

On TopoQA, DeepSeek-V3 has the highest reported accuracy and shortest response time, while Qwen2.5-72B uses fewer tokens with competitive accuracy. The authors describe Qwen2.5-72B’s 94.55% accuracy as near-state-of-the-art despite lower model scale.

3.3 Benchmark on TopoQA dataset

  • Qwen3 uses 722,308 total tokens and takes 79.76 seconds on average despite a median token count of 3,710.5.The authors attribute the inefficiency to over-reasoning, especially redundant tool-call chains in combinatorial queries that can cause token overflow and reveal task-routing instability.Source: §3.3 · p. 8
  • Qwen2.5-72B’s 94.55% accuracy versus DeepSeek-V3’s 96.40% is presented as evidence that model scale is not decisive in this specialized domain.The authors call Qwen2.5-72B optimal under strict computational constraints when resource efficiency outweighs marginal accuracy gains.Source: §3.3 · p. 8
  • DeepSeek-V3 achieves the highest accuracy (96.40%) and shortest average response time (38.51 seconds), while Qwen2.5-72B uses 566,258 total tokens at 94.55% accuracy.Table 2 compares model performance on TopoQA; Qwen2.5-72B has superior computational economy.Source: Table 2 · p. 8, §3.3 · p. 8

3.4 Benchmark on TopoOQ dataset

TopoOQ evaluations show model-specific quality trade-offs and a balanced efficiency profile for Qwen2.5-72B, while response speed and token use vary substantially across models.

3.4 Benchmark on TopoOQ dataset

  • Third-party TopoOQ evaluations score the system from 7.2 to 7.9 across accuracy, completeness, and consistency dimensions.The benchmark reports effective performance across Qwen2.5-72B, Qwen3, and DeepSeek-V3.Source: §3.4 · p. 8, Table 3 · p. 9
  • Qwen2.5-72B scores 7.25 versus 7.13–7.20 for accuracy, while Qwen3 reaches 7.39 completeness and 7.94 factual consistency in open-ended analysis.Qwen2.5-72B is described as optimal for precise property retrieval, while Qwen3 excels in complex relational reasoning.Source: §3.4 · p. 8
  • DeepSeek-V3 records 91.34s average time and 1.35M tokens, while Qwen3 records 207.06s and 1.96M tokens amid quality and latency trade-offs.The authors attribute Qwen3’s overhead to over-reasoning during multi-agent tool selection.Source: §3.4 · p. 8
  • Qwen2.5-72B operates at 1.46× the speed of DeepSeek-V3 while delivering 95% of Qwen3’s analytical depth, with a composite score of 7.46.The comparison presents Qwen2.5-72B as a balanced solution with reduced computational demands.Source: §3.4 · p. 8, Table 3 · p. 9

3.5 Case study

The case study links material retrieval and constrained crystal generation to first-principles verification of SrSbO3, then stores the resulting classification for reuse.

3.5.1 Case 1: Material retrieval & reasoning workflow

  • The retrieval workflow returns Sb2S3’s 1.2866 eV band gap and 56.359 total dielectric constant, then uses knowledge-graph templates for topological classification queries.For the classification query, the KG Agent retrieves five relevant Cypher templates and returns structured data for the Core Agent to synthesize.Source: §3.5.1 · p. 9
  • Under space-group R-3m and Sr, Sb, and O constraints, CrystalFormer generates Sr2Sb2O7 and SrSbO3 for 3D visualization.The Core Agent identifies design requirements and engages the MG Agent.Source: §3.5.1 · p. 9

3.5.2 Case 2: Topological Property calculation

  • For SrSbO3 verification, the Core Agent delegates to the CP Agent, which confirms requirements before a three-stage plan covering input preparation, DFT, and topological analysis.The workflow uses LangGraph state tracking and automated error recovery.Source: §3.5.2 · p. 9
  • The DP Agent prepares VASP POSCAR, INCAR, and KPOINTS files from SrSbO3’s structure, and the RE Agent validates the package before calculation.Source: §3.5.2 · p. 9
  • The workflow suspends while HPC jobs run and polls their status; when an SCF calculation fails to terminate, it rolls back and resubmits until outputs validate.The workflow preserves state in MongoDB and tracks jobs asynchronously; the recovery cycle is demonstrated by intentionally removing OUTCAR termination markers.Source: §3.5.2 · p. 9
  • DFT-based topological-invariant analysis with SymTopo confirms SrSbO3 as a topological crystalline insulator.The RP Agent parses converged wavefunction outputs and classifies the material’s topological phase.Source: §3.5.2 · p. 12
  • The RP Agent stores structured composition, electronic-property, and classification results in Neo4j, allowing later knowledge-graph queries to retrieve SrSbO3’s classification without recalculation.Source: §3.5.2 · p. 12

4 Conclusions

TopoMAS integrates material databases, crystal generators, online resources, and first-principles engines into a query-to-update workflow. The authors report SrSbO3’s discovery and benchmark performance for Qwen2.5-72B, while identifying adaptive routing and broader computational coverage as future work.

4 Conclusions

  • TopoMAS integrates multi-source databases, crystal generators, online resources, and first-principles engines in a query–reasoning–calculation–update workflow that identified SrSbO3 as a new topological crystalline insulator.Source: §4 · p. 12
  • Qwen2.5-72B achieves over 94% accuracy on structured tasks and a 7.46/10 composite score for open-ended inquiries, using 74–83% fewer tokens and responding twice as fast.Source: §4 · p. 12
  • Future work will develop adaptive computation routing, extend the framework beyond topological properties, and strengthen fault tolerance.Source: §4 · p. 12

Conflicts of interest

The authors declare no conflicts of interest.

Conflicts of interest

Notes and references

The section consists of bibliography entries numbered 1–61, presented across three passages.

Notes and references

Loading 2507.04053v1…