Source-linked AI summary

QUASAR: A Universal Autonomous System for Atomistic Simulation and a Benchmark of Its Capabilities

Fengxu Yang, Jack D. Evans

arXiv:2602.00185v2cond-mat.mtrl-scics.AI

TL;DR

Computational-chemistry agents are limited by rigid, human-crafted scaffolding and narrowly scoped tools. QUASAR introduces a universal atomistic system that combines adaptive planning, multiscale workflow orchestration, iterative refinement, and hybrid knowledge retrieval. Across tiered benchmarks, it completed sophisticated routine and frontier tasks, supporting its potential as a general atomistic reasoning system while retaining important limitations tied to LLM knowledge and reproducibility.

  • Problem

    Existing computational-chemistry agents rely on rigid human-crafted scaffolding and specialized functions whose scope and complexity limit adaptability.

  • Method

    QUASAR uses coordinated Strategist, Operator, and Evaluator agents with adaptive planning, iterative refinement, and hybrid knowledge retrieval for multiscale atomistic workflows.

  • Results

    QUASAR executed sophisticated computational-chemistry tasks across Tier III without auto-improvement, including end-to-end screening and method selection for frontier cases.

  • Takeaways & Limitations

    The findings provide initial evidence that agentic AI can support computational-chemistry workflows by handling routine orchestration while humans provide direction, judgment, and validation.

  • Takeaways & Limitations

    Performance remains bounded by the selected LLM’s domain knowledge and in-context reasoning, and LLM errors can produce plausible but physically unsound outputs.

Abstract

from arXiv · show

The integration of large language models (LLMs) into materials science offers a transformative opportunity to streamline computational workflows, yet current agentic systems remain constrained by rigid, carefully crafted domain-specific tool-calling paradigms and narrowly scoped agents. In this work, we introduce QUASAR, a universal autonomous system for atomistic simulation designed to facilitate production-grade scientific discovery. QUASAR autonomously orchestrates complex multi-scale workflows across diverse methods, including density functional theory, machine learning potentials, molecular dynamics, and Monte Carlo simulations. The system incorporates robust mechanisms for adaptive planning, context-efficient memory management, and hybrid knowledge retrieval to navigate real-world research scenarios without human intervention. We benchmark QUASAR against a series of three-tiered tasks, progressing from routine tasks to frontier research challenges such as photocatalyst screening and novel material assessment. These results suggest that QUASAR can function as a general atomistic reasoning system rather than a task-specific automation framework. They also provide initial evidence supporting the potential deployment of agentic AI as a component of computational chemistry research workflows, while identifying areas requiring further development.

1 Introduction

Computational chemistry is well suited to LLM-based automation, but existing agentic systems remain constrained by rigid scaffolding and narrowly scoped tools. QUASAR addresses these limitations with a universal atomistic computation system intended to coordinate diverse workflows.

  • LLM-based automation can streamline simulation setup, data interpretation, and error handling while reducing specialized software burdens and continuous human oversight.
  • Existing computational-chemistry agents rely heavily on human-crafted tools, fine-grained decompositions, and rigid workflows that risk underusing improving LLM capabilities.
  • Human-defined functions can impose scope risks because they may behave inappropriately outside anticipated use cases, whereas LLMs can adapt to edge cases and unconventional workflows.
  • Specialized multi-agent designs increase coordination complexity and engineering burdens, undermining scalability, reliability, and adaptability in real-world settings.
  • QUASAR is introduced as a production-level universal atomistic computation system that coordinates workflows from quantum-mechanical calculations to classical molecular simulations.

2 System Architecture and Methods

QUASAR combines a three-agent architecture with adaptive planning, context management, persistent execution, and hierarchical knowledge retrieval. Its design coordinates multiscale simulations while supporting iterative refinement and robust operation across computational environments.

  • Architecture: The Strategist decomposes research objectives, the Operator executes simulation tasks, and the Evaluator assesses results and returns unsatisfactory outputs for refinement.
  • Multiscale integration: QUASAR integrates DFT, machine-learned potentials, molecular dynamics, and Monte Carlo software through common structure-manipulation and analysis tools.
  • Adaptive planning: A double-pass planning mechanism reviews initial plans for omitted prerequisites before execution, reducing the risk of suboptimal or physically inconsistent workflows.
  • Adaptive planning: Optional automated improvement cycles and user-adjustable granularity and accuracy parameters support iterative optimization while balancing decomposition, computational cost, and precision.
  • Context management: Context sharing is optimized across tasks and runs by condensing verified actions and outcomes while discarding noise and failed attempts.
  • Execution robustness: Persistent checkpointing enables recovery from interruptions, while restart-aware execution can resume simulations from intermediate states rather than repeating completed work.
  • Knowledge retrieval: Hierarchical retrieval escalates from internal knowledge to documentation, repository inference, and finally external resources; disabling external retrieval caused consistent RASPA3 input-file failures.
  • Deployment: QUASAR supports HPC execution through containerization and batch-compatible operation, but does not currently provide generalized agent-driven HPC job submission.

3 Results

QUASAR was evaluated with a three-tier benchmark spanning routine single-step calculations, multi-step workflows, and open research challenges. The benchmark progresses from established computational procedures to photocatalyst, gas-separation, and novel-material assessments.

  • Benchmark design: The benchmark used three tiers with repeated executions; Tier I and II outcomes were based on three runs, while Tier III cases were executed twice.
  • Tier I. Task Execution: Tier I tested single-step execution across DFT k-point convergence, molecular-dynamics equilibration, and helium void-fraction analysis.
  • Tier II. Multi-step Workflows: Tier II evaluated decomposition of scientific questions into workflows for NiO band gaps, UiO-66 CO2 adsorption isotherms, and aluminum melting points.
  • Tier III. Frontier Challenges: Tier III targeted recent or open problems, including La-doped perovskite photocatalyst screening, Xe/Kr-selective COF identification, and prediction for a newly generated material.

4 Discussion

QUASAR achieved accurate, rigorous performance across Tier I and Tier II, while Tier III demonstrated end-to-end execution of novel workflows with minimal or no human intervention. The discussion also identifies reproducibility, model dependence, and the continuing need for human oversight as important boundaries.

  • QUASAR achieved accurate results with efficiency and rigor across all Tier I and Tier II cases.
  • After an inaccurate first NiO band-gap run, one or two auto-improvement contexts led QUASAR to switch from DFT+U to higher-fidelity HSE.The authors attribute the initial choice to strong training priors favoring DFT+U for NiO.
  • Tier III photocatalyst screening recreated the published result, while gas-separation and virtual-MOF cases were physically reasonable and methodologically reliable.
  • Without auto-improvement, QUASAR completed Tier III screening, method selection, and analysis end-to-end, suggesting minimal or no human intervention for complex workflows.
  • Repeated trials did not undermine the overall user-request goal, and recorded inputs plus step-by-step explanations support reproducing prior trajectories.
  • QUASAR’s generalization beyond the evaluated benchmark suite requires further study, while performance remains bounded by the selected model’s domain knowledge and in-context reasoning.Errors may appear physically or numerically unsound only under expert scrutiny.
  • Human oversight remains necessary to detect agent-introduced errors and ensure outputs are physically plausible.

5 Summary and Outlook

QUASAR is presented as an open-source, portable, universal atomistic computation system aimed at production-grade scientific discovery. Tier III results support human–AI collaborative workflows, while broader benchmarking and changing model choices remain necessary for evaluation.

  • QUASAR is an open-source, portable, universal atomistic computation system designed to advance agentic AI toward production-grade scientific discovery.
  • Tier III benchmarking found that properly orchestrated LLMs executed sophisticated computational chemistry tasks with scientific rigor comparable to human researchers.
  • QUASAR is positioned as a human–AI collaborative workflow in which AI handles routine orchestration while humans provide scientific direction, judgment on ambiguous cases, and validation.
  • Selecting the best-performing model remains challenging because frontier and specialized computational-chemistry LLMs continue to emerge.
  • Future benchmarks should extend beyond the limited evaluated set to assess scientific success, underlying capabilities, strategy quality, and decision-making efficiency.
  • As execution, debugging, and workflow construction become automated, human researchers may shift toward conceptual framing and theoretical innovation.
Loading 2602.00185v2…