SQD-Agent: LLM-driven agentic framework for Quantum Chemistry workflows

I. INTRODUCTION

• SQD Agent targets the expertise and workflow-assembly burden of quantum chemistry, where users must configure chemistry, algorithms, circuits, hardware execution, mitigation, and classical post-processing across software libraries.Existing libraries provide computational building blocks, but users must still assemble them into a reproducible, hardware-aware workflow,.Source: §I · p. 1
• The framework separates LLM-based planning from specialized agents and library-backed execution for chemistry setup, SQD-family calculations, hardware submission, mitigation comparison, profiling, debugging, and reporting.The execution layer uses PySCF, Qiskit, ffsim, and qiskit-addon-sqd.Source: §I · pp. 2–3
• The framework is modular, with integration points for additional SQD-family algorithms, tools, and profiling or benchmarking components.Examples include extended SQD, SQdrift [4], and Fulqrum,.Source: §I · p. 3
• Its hardware agent supports circuit optimization, QPU execution, job retrieval, mitigation suggestions, and post-run comparison of accuracy–cost trade-offs.Cost includes shots, QPU jobs, runtime, and classical post-processing.Source: §I · p. 3
• The evaluation covers six molecular systems on simulators and IBM quantum systems, demonstrating end-to-end agent-mediated execution and result dashboards.Source: §I · p. 3
• SQD Agent is positioned as a quantum-native execution and analysis layer, managing circuit generation, hardware-aware transpilation, sampling, mitigation, SQD diagonalization, and performance reporting rather than replacing SQD algorithms.Source: §I · p. 3

A. Agent-Callable SQD Execution Engine

• The execution engine accepts validated workflow requests and returns numerical outputs, diagnostic traces, and reproducibility metadata through conversational or scripted interfaces.Agent decisions pass through the structured SQDRunSpec contract, which prevents free-form code execution and makes runs auditable.Source: A. Agent-Callable SQD Execution Engine · p. 3
• Chemistry preparation builds Hartree–Fock references and molecular integrals, optionally computes reference energies, and selects active spaces using supported orbital-selection policies.A capacity-preserving refit prevents silent electron loss during active-space truncation.Source: A. Agent-Callable SQD Execution Engine · p. 3
• The SQD loop samples circuits, recovers target-electron-number configurations, batches selected subspaces, then projects and diagonalizes the molecular Hamiltonian.Merged batches pool configurations and may improve accuracy; independent batches reduce diagonalization size and expose parallelism.Source: A. Agent-Callable SQD Execution Engine · p. 4
• Backend abstraction converts simulator and hardware samples to a common bitstring-count interface, while the modular execution layer can accommodate additional backend adapters.The architecture does not assume IBM-specific hardware at the SQD level.Source: A. Agent-Callable SQD Execution Engine · p. 4
• Hardware mitigation is configurable: agents can suggest compatible strategies before execution and compare measured time–accuracy trade-offs after runs.The report records settings including backend, shot count, QPU time, classical diagonalization time, recovered energy, and reference deviation where available.Source: A. Agent-Callable SQD Execution Engine · p. 4
• Resource logging and on-demand profiling produce timing summaries, circuit metrics, convergence traces, and artifacts for later reporting without rerunning calculations.The design describes its timing breakdown as lightweight and reproducible, not a full low-level performance model.Source: A. Agent-Callable SQD Execution Engine · p. 4
• A cache-first policy reuses compatible stored results, while saved hardware job identifiers support retrieval of interrupted runs.Run signatures include computation settings and software environment.Source: A. Agent-Callable SQD Execution Engine · pp. 4–5
• The deterministic execution layer is extensible to new SQD-family algorithms, classical solvers, and quantum backends through modular interfaces.Source: A. Agent-Callable SQD Execution Engine · pp. 3, 5

B. Agentic Orchestration and Workflow Control

• The orchestrator routes interpreted tasks to specialists, which construct validated structured requests; the execution layer returns task-specific results and artifacts rather than running arbitrary LLM-generated code.Source: B. Agentic Orchestration and Workflow… · p. 5
• Specialists divide simulator execution, hardware operations, profiling, and reporting, with practical cost accounting spanning QPU resources and classical post-processing.Narrow operational scopes are intended to make workflows auditable and prevent silent changes to scientific or hardware parameters.Source: B. Agentic Orchestration and Workflow… · p. 5
• Human-in-the-loop confirmation keeps scientific oversight with the researcher for molecular geometry and requires explicit confirmation of key hardware-run choices before QPU submission.The user confirms geometry when a molecular name rather than a complete geometry is supplied, and confirms backend, ansatz, shot count, and mitigation strategy for hardware execution.Source: B. Agentic Orchestration and Workflow… · p. 5
• The cache-first policy reuses compatible artifacts and retrieves completed results for interrupted hardware jobs, with task-specific signatures preserving traceability.Profiling signatures also record machine and software configuration because performance measurements depend on the system used.Source: B. Agentic Orchestration and Workflow… · p. 5
• New ansätze, SQD variants, solvers, or backends can be exposed through task-specific specifications and validation policies without changing the agent layer.Source: B. Agentic Orchestration and Workflow… · p. 5
• SQD Agent is distinguished from general chemistry agents by its quantum-native workflow, including hardware execution, mitigation comparison, and quantum–classical performance reporting.Source: B. Agentic Orchestration and Workflow… · p. 6

III. EXPERIMENTAL EVALUATION OF AGENTIC SQD WORKFLOWS

• The evaluation asks whether natural-language prompts can be converted into valid SQD-family workflows and whether simulator runs recover reference energies.Source: §III · p. 6
• It examines whether the framework exposes quantum and classical resource costs and supports reproducible reporting from cached artifacts, convergence traces, and timing summaries.Source: §III · p. 6
• A final evaluation question tests whether the hardware agent can execute and compare mitigation strategies on a real QPU.Source: §III · p. 6

• PySCF, Qiskit and Qiskit Aer, ffsim, and qiskit-addon-sqd support electronic-structure preparation, circuit construction and simulation, fermionic state preparation, and sample-based diagonalization.Source: A. Evaluation Protocol and Setup · p. 6
• The hardware experiment used Qiskit Runtime on IBM Quantum backend ibm_kingston, identified as a 156-qubit backend.Source: A. Evaluation Protocol and Setup · p. 6
• Simulator benchmarks used STO-3G and compared HF, UCJ, LUCJ, and fixed hardware-efficient ansätze; unless stated otherwise, they used 300,000 shots per circuit and up to eight iterations.The N2/LUCJ hardware experiment tested submission, retrieval, SQD post-processing, and mitigation comparison; saved runs include specifications, environment, circuit metrics, convergence, timing, and artifacts.Source: A. Evaluation Protocol and Setup · p. 6

• The request to run SQD energy estimation for formaldehyde in STO-3G and compare ansätze was classified as simulator benchmarking and dispatched to the SQD-family execution agent.The output included an ansatz-comparison table, circuit metrics, convergence traces, and saved artifacts; Table II summarizes SQD Agent benchmarks in STO-3G on a Qiskit simulator, with detailed results in Appendix A.Source: B. Natural-Language to SQD Workflow… · p. 6, Table II · p. 7
• The reporting agent generated convergence plots and summary tables from the same saved artifacts without requiring manual reconstruction of the command-line invocation.Source: B. Natural-Language to SQD Workflow… · pp. 6–7
• The generated execution specification and artifacts were saved with the result, supporting reproducibility of the run.Source: B. Natural-Language to SQD Workflow… · p. 7

• Correlated UCJ/LUCJ-style ansätze recover reference energies with small deviations in tested full-space calculations, while active-space runs expose truncation error when the active window omits correlation-relevant orbitals.Source: C. Simulator Accuracy and Resource… · p. 7
• UCJ and LUCJ provide chemically accurate or near-chemical-accuracy estimates in tested full-space cases, whereas hardware-efficient circuits have lower structural complexity but do not consistently provide the same accuracy across molecules.Source: C. Simulator Accuracy and Resource… · p. 7
• Recorded metrics include qubit count, transpiled depth, gate counts, shots, runtime, and SQD convergence traces, enabling ansatz comparisons by final energy and execution cost.Source: C. Simulator Accuracy and Resource… · p. 7

• Profiling reports break down sampling, SQD post-processing, subspace diagonalization, and serialization timing, interpreted relative to the recorded system configuration.Timing interpretation depends on processor, memory, operating system, Python environment, library versions, and linear-algebra thread settings.Source: D. Profiling, Caching, and Convergence… · p. 7
• For repeated simulator or hardware tasks, a compatible cached artifact is summarized instead of rerunning the workflow; saved hardware job identifiers also allow retrieval after interruptions.This can avoid additional queue time, shots, and backend access for hardware experiments.Source: D. Profiling, Caching, and Convergence… · p. 7
• Saved per-iteration SQD traces support energy-versus-iteration plots, subspace-size summaries, and ansatz-comparison tables.Source: D. Profiling, Caching, and Convergence… · p. 7

• The N2/LUCJ workflow ran on ibm_kingston with 100 000 shots per job, using a 20-qubit logical circuit transpiled to 156 physical qubits.The selected hardware circuit had total depth 210 and two-qubit gate depth 42.Source: E. Hardware Execution and Mitigation… · p. 7
• Within 10−3 mHa of FCI, all tested mitigation settings recovered the reference, while the unmitigated baseline had the lowest observed runtime.The comparison is for N2/LUCJ on ibm_kingston; Table III reports hardware mitigation results against FCI.Source: E. Hardware Execution and Mitigation… · p. 7, Table III · p. 8
• The hardware experiment supports claims that SQD Agent translates natural-language requests into SQD workflows and supports simulator and hardware execution paths.Resource metrics and convergence traces are saved, cached artifacts can be reused, and hardware-specific reports support mitigation suggestion or comparison.Source: E. Hardware Execution and Mitigation… · p. 8
• The hardware result is a capability validation, not a broad chemistry benchmark, because it covers one molecule, one ansatz, and one backend.The simulator benchmarks are also limited to small molecules in a minimal basis with available FCI or CASCI references.Source: E. Hardware Execution and Mitigation… · p. 8

Evaluation

• On a 156-qubit IBM processor, N2 SQD reproduced the exact FCI energy within chemical accuracy; simulator evaluations used CASCI and CCSD references.The work describes this as execution validation, not a claim of quantum advantage or a new SQD algorithm.Source: §IV · p. 8

Implementation

• Modular Python components support extension to new molecular systems, while cache-first execution and timestamped artifacts support reproducibility.Artifacts include commands, metadata, traces, metrics, and outputs.Source: §IV · p. 9

Future work

• The framework is intended to scale toward larger chemical systems, but that portability does not change the underlying SQD algorithms’ computational scaling or accuracy properties.Transition-metal clusters, catalytic centers, and biomolecular systems are named as future targets.Source: §IV · p. 9

Scope and limitations

• Experiments are limited to small molecules in a minimal STO-3G basis; the hardware demonstration is one N2 run on one backend, with no run-to-run error bars.Reported energies come from single runs with an unseeded sampler; cached artifacts support reproducibility rather than bitwise-identical recomputation.Source: §IV · p. 9
• The agent is an orchestration and policy framework, not an autonomous reasoner, and workflow reconfiguration remains human-supervised rather than closed-loop.Hardware submission and fetched geometries require explicit human confirmation; validated execution feedback is planned to guide later workflow choices.Source: §IV · p. 9

Future work

• The current framework supports one algorithm family, SQD, for molecular ground-state estimation; broader task and algorithm support is planned.A dedicated pilot user study is also planned because preliminary use with a small expert set cannot support statistically meaningful workflow-efficiency conclusions.Source: §IV · p. 9

Benchmark metrics

• Benchmark entries report energies in Hartree, deviations from full-space FCI in millihartree, runtimes in seconds, and circuit metrics including qubits, depth, and gate counts.“Full” denotes the full orbital space, while “active” denotes the selected active-space window; a dash means a metric is inapplicable or unrecorded.Source: Appendix A · p. 10

H2O benchmark

• The H2O benchmark compares full-space and active-space SQD variants using ncore = 2, ncas = 5, and nelecas = (3, 3).The caption identifies the active-space window but provides no benchmark values.Source: Table IV · p. 10

LiH benchmark

• The LiH benchmark compares full-space and active-space SQD variants, with the selected active space spanning all six orbitals.The caption provides no benchmark values.Source: Table V · p. 11

N2 benchmark

• The N2 benchmark compares full-space and active-space SQD variants using ncore = 4, ncas = 6, and nelecas = (3, 3).The caption identifies the active-space window but provides no benchmark values.Source: Table VI · p. 11

CH2O benchmark

• The appendix also includes a CH2O benchmark table reporting SQD Agent results.The caption does not specify a full-space or active-space comparison.Source: Table VII · p. 11