Quantum-HPC Workflows Across Multiple Quantum Computing Platforms: Two Case Studies

AI summary

Overview

Research area: Hybrid quantum–high-performance computing (quantum-HPC), specifically workflow management for orchestrating classical supercomputers and multiple quantum processing units (QPUs) in a single application.

Technical level: Intermediate. The paper is written as a systems/infrastructure contribution rather than a quantum-algorithm paper, so readers need some familiarity with quantum circuits (VQE, QPE) and HPC job scheduling, but the emphasis is on orchestration and software architecture.

Scope (one sentence): The paper describes and demonstrates a Tierkreis-based workflow system that connects the Fugaku supercomputer to the Quantinuum Reimei and ibm_kobe quantum computers, through two end-to-end case studies in quantum chemistry and ground-state energy estimation.

What This Paper Is About

Practical quantum algorithms always need classical computation around them (problem encoding, circuit optimisation, post-processing), and HPC systems can supply that classical power — but today each quantum-HPC application is a bespoke assembly of schedulers, scripts and glue code tied to one facility. The authors adapt the Tierkreis quantum workflow management system to the HPC environment and show it coordinating Fugaku with two different QPUs. The goal is not a new algorithm or a micro-benchmark, but a reusable, hardware-agnostic integration layer that shifts the burden of integration from the scientist to the infrastructure.

Key Contributions

1. HPC extensions to Tierkreis. The authors implement pjsub (the Fujitsu TCS job submission command) as a Tierkreis executor, add quantum submission through sqcsub via a shell executor plus a custom worker, and introduce reusable worker interfaces for HPC and quantum resources.

2. A hybrid biomolecular excited-state workflow. A subtractive ONIOM workflow combining CASCI (classical, via NTChem on Fugaku) with time-evolved quantum selected configuration interaction (TE-QSCI), parallelising quantum circuit generation, compilation and submission across target eigenstates.

3. The first end-to-end multi-modality quantum-HPC workflow (to the authors' knowledge). A single workflow in which Fugaku coordinates VQE on the 156-qubit superconducting ibm_kobe and QPE on the 20-qubit trapped-ion Reimei, exploiting the speed of one device and the accuracy of the other.

4. A demonstrator of workflow portability. Switching backends — for example from Reimei to the reimei-simulator, or between ibm_kobe, the reimei-simulator and Qiskit Aer — requires only a small configuration change or a single line, because task definitions are separated from environment-specific implementations.

Main Findings

• Workflow is dominated by the classical chemistry kernel. In the biomolecular case (evolution time 7.5), total workflow execution time was 3240 s, 4066 s and 3479 s for the three backends reported, with the NTChem task taking 2653 s, 2651 s and 2689 s — approximately 80% of total execution time.

• Circuit preparation is a non-trivial cost. "Circuit Prep" took 141.8 s, 139.0 s and 143.3 s across the three runs, which the authors say suggests future optimisation and parallelisation of this stage is necessary. Reported "Circuit Exec" values were 0.5726 s, 240.8 s and 21.9 s.

• Queue time was small but variable. NTChem queue waiting times were 29 s, 53 s and 98 s; the authors note this depends on supercomputer workload and congestion, and that a relatively small number of nodes were used.

• Excitation energies reproduced across backends. For the T0 and S1 states, TE-QSCI in an (8e,8o) active space gave 1.29/1.26 eV (ibm_kobe), 1.26/1.23 eV (reimei-simulator) and 1.28/1.30 eV (Aer), against CASCI reference values of 1.26/1.29 eV in (8e,8o) and 1.44/1.48 eV in (6e,6o). The experiments omitted error correction and used the reimei-simulator because of a scheduled hardware upgrade, so results may differ slightly from the original reference work.

• Warm-starting QPE with VQE improves accuracy. On the 4×4 transverse-field Ising model with h = 3J, the exact ground state is E_exact = −51.45 J and the workflow's VQE state gives E_VQE = −48.00 J. The final hardware estimate was (E − E_exact)/J = 0.61 ± 0.36 after 300 QPE shots, versus a noiseless VQE+QPE value of 0.25 ± 0.21 and VQE alone at 2.47 ± 0.10 on hardware.

• A classically optimised product state is a much weaker QPE input. Product + QPE gave 0.8 ± 1.1 in noiseless simulation (no hardware result reported), compared with 0.25 ± 0.21 for VQE + QPE, confirming that a better initial state matters more than the QPE stage alone.

• The hardware gap is attributed to compounding error. VQE + QPE remains clearly better than VQE alone on hardware, but the advantage is slightly smaller than in simulation; the authors believe this comes from processing an imperfectly optimised initial state on a noisy device.

• Portability is demonstrated concretely. The biomolecular workflow ran unchanged in structure on a local simulator, the reimei-simulator and ibm_kobe, and the second workflow dispatched to two QPU modalities without changing the workflow definition.

Methodology in Plain English

The authors treat a hybrid application as a graph of independent tasks rather than a monolithic script. Each task has a typed interface declaring its inputs and outputs, so mistakes are caught when the workflow is built rather than at runtime. Tasks are supplied by libraries called workers; executors decide where each task actually runs — a local process, a batch job on Fugaku via pjsub, or a quantum circuit sent through sqcsub to a QPU. All intermediate data and metadata are checkpointed to the shared filesystem, so a workflow can be paused, resumed or restarted after an error without losing completed work.

For the chemistry case, they combine a classical HF+CASCI calculation in a (6e,6o) active space on 64 Fugaku nodes (4 MPI processes × 12 OpenMP threads per node) with TE-QSCI circuits on 16 qubits, expanding to an (8e,8o) active space, then merging sampled configurations and diagonalising the subspace on Fugaku. Shots were 200 for Qiskit Aer and the reimei-simulator and 20,000 for ibm_kobe, reflecting differing access costs, latency and fidelities.

For the second case, they split the problem by device strength. VQE — which needs many rounds and many shots — runs on the faster ibm_kobe using a symmetry-respecting layered ansatz optimised with SPSA (four measurements per iteration, 500 iterations, 500 shots per X/Z measurement). Because a converged VQE yields a classical description of the state through its ansatz and optimal parameters, that state can be re-prepared on Reimei without any quantum communication. Reimei then runs information-theory QPE with one ancilla qubit, k_max = 20, λ = 0.01π/J, 300 shots, plus circuit compression, coherent noise suppression and leakage-detection gadgets on the three remaining qubits.

Why This Matters

Impact on research. The paper offers a blueprint for making quantum-HPC applications portable: application logic stays agnostic to hardware, while integration effort moves into executors and workers maintained at the infrastructure level. It also provides a concrete existence proof that a single workflow can span two QPU modalities, which matters as heterogeneous quantum resources become normal rather than exceptional.

Real-world applications:

• Photochemistry and biomolecular science, where excited states govern how proteins and DNA interact with light.
• Quantum chemistry workflows such as QSCI and SQD that mix quantum sampling with distributed classical diagonalisation.
• Materials and condensed-matter problems, illustrated by ground-state energy estimation for the transverse-field Ising model.
• Hybrid methods that extend the same pattern, such as QM/MM and quantum machine learning, which the paper cites as adjacent use cases.

Industry relevance. The work sits directly at the integration layer that vendors, HPC centres and cloud providers must agree on. It shows facility-specific quirks (Fujitsu's scheduler, gRPC access, direct low-latency links, Qiskit backends) can be absorbed by a workflow system instead of being re-implemented by every application team, and it demonstrates a practical speed-versus-accuracy division of labour between superconducting and trapped-ion hardware.

Future Directions

• Optimise and parallelise circuit preparation, which the timings show is a non-negligible cost and which the authors explicitly flag as needing future work.
• Exploit concurrency more aggressively. The map operation over circuit lists could in principle dispatch tasks concurrently across HPC nodes; in these experiments all such tasks were run locally as that was judged overkill.
• Re-run on real Reimei hardware with error correction. The chemistry case used the reimei-simulator and omitted the error correction of the original work, leaving open how the workflow performs under the full hardware and error-correction stack.
• Standardise the quantum-HPC software stack. The paper notes the community has not consolidated around a tool or stack, and positions its executor/worker design as a candidate pattern for further integration, including lower-level interfaces such as QDMI and QRMI.

Target Audience

Infrastructure and research software engineers building hybrid quantum-classical systems; HPC centre staff supporting quantum resources; quantum algorithm developers whose workflows are growing beyond a single device; and computational chemists interested in how TE-QSCI-style pipelines are orchestrated in practice. Readers seeking new quantum algorithms, detailed chemistry results or performance micro-benchmarks should consult the referenced chemistry work instead, as the paper explicitly scopes itself to systems integration.

Authors’ abstract

We describe the Quantum-HPC hybrid system comprising the supercomputer Fugaku, the QPUs Quantinuum Reimei and ibm_kobe, and the Tierkreis hybrid workflow software. We present two demonstrations of the system at work. In the first, we compute the excited states of a biomolecular system using Fugaku and Reimei (simulator) to handle classical and quantum parts of the workflow respectively. In the second, Fugaku coordinates Reimei and ibm_kobe working together in a combined VQE + QPE workflow to compute the ground state energy of a simple system.