arXiv:2609.39302
SQD-Agent: LLM-driven agentic framework for Quantum Chemistry workflows
Kislaya Tiwari, Anupama Ray
Abstract
Quantum algorithms and quantum hardware are advancing towards a promising paradigm for scientific applications. However, translating domain-specific problems into executable hybrid quantum-classical workflows remains a significant barrier for application researchers due to the required expertise in quantum algorithms, nuances in quantum programming, and hardware-aware system integration. At the same time, AI and primarily LLM based agents are increasingly capable of interpreting natural-language intent, reasoning over complex workflows, and translating high-level objectives into executable code and building computational pipelines. In this work, we introduce SQD Agent, an LLM-based agentic framework that translates natural-language user intent into executable workflows for Quantum Chemistry applications where algorithms from the Sample-Based Quantum Diagonalization (SQD) family are used. By automating this translation, SQD Agent reduces the level of human expertise and configuration overhead required, thereby simplifying experimentation in hybrid quantum-classical settings for application researchers new to quantum. SQD Agent adopts a modular and extensible architecture that supports seamless integration of heterogeneous quantum backends, classical solvers, and workflow components, ensuring adaptability to rapidly evolving quantum ecosystems. The framework further incorporates interactive capabilities for on-demand profiling, bottleneck analysis, resource optimization, intelligent result caching, and convergence visualization. Key features include quantum chemistry experiments, error mitigation on real quantum hardware, together with analysis of candidate mitigation schemes in terms of their potential error-recovery behavior and computational budget, helping users understand their practical trade-offs and decide which strategies to explore in subsequent experiments.
AI Summary
SQD-Agent addresses the software-integration burden of running sample-based quantum diagonalization (SQD) chemistry calculations across classical resources, simulators, and quantum hardware. It provides an LLM-driven multi-agent interface that translates natural-language requests into structured, validated workflow specifications rather than executing arbitrary generated code. Specialist agents coordinate chemistry preparation, circuit construction, sampling, SQD post-processing, hardware submission, mitigation comparisons, profiling, and reporting using PySCF, Qiskit, `ffsim`, and `qiskit-addon-sqd`. Cached artifacts and execution metadata support retrieval and reporting; geometries and hardware-resource choices require human confirmation. Evaluation consists of example prompt-to-workflow execution, small-molecule STO-3G simulator benchmarks, and an $\mathrm{N}_2$ hardware experiment. The latter uses a 20-logical-qubit LUCJ circuit on the 156-qubit `ibm_kingston` backend, with 298 transpiled two-qubit gates, two-qubit depth 42, and 100,000 shots per mitigation setting. The hardware table reports absolute FCI deviations of approximately $0.0195$ mHa without mitigation, $0.00180$ mHa with dynamical decoupling, and $0.12$–$0.13$ mHa with twirling-based settings. These values differ from the accompanying claim that every setting achieves below $10^{-3}$ mHa; simulator summary and appendix values also differ. The contribution is workflow orchestration, not a new SQD algorithm or quantum-advantage demonstration. Stated limitations include minimal-basis systems with tractable classical references, one hardware molecule/backend, single-run energies without error bars, machine-dependent profiling, LLM-dependent reliability, and human-supervised rather than closed-loop adaptation. No statistically meaningful comparison of workflow efficiency against manual execution is provided; a user study is proposed.
Error found: The claimed reduction in required expertise and configuration overhead is not quantitatively evaluated against manual execution, and the paper explicitly defers such testing to future work. The abstract's mitigation recovery-and-budget guidance is only partially supported by compatibility suggestions and retrospective comparisons, while prospective recovery-range and compute-budget recommendations remain under development. The simulator summary contradicts the appendix on LiH and N2 energy errors and on which ansatz achieves the best listed LiH and H2O energies. The hardware paragraph incorrectly claims every mitigation setting reaches FCI within 0.001 mHa, whereas the displayed energies imply errors of approximately 0.00185–0.13025 mHa. The unmitigated hardware run is incorrectly called the lowest observed runtime despite having the highest reported total; that ranking holds only after subtracting its disclosed queue delay.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Accepted in IEEE International Conference on High Performance Computing (HiPC 2026)