Demonstration of Parallel Multi-QPU Execution for Fragment-Based Quantum Chemistry Using On-Premises Hardware

AI summary

Overview

Research area: Quantum computing for quantum chemistry — specifically fragment-based electronic structure (FMO) executed across multiple small quantum processing units (QPUs) installed in an HPC facility.

Technical level: Advanced. The paper assumes familiarity with second-quantized electronic structure, selected configuration interaction, and quantum circuit sampling, though the core argument (parallelism increases throughput) is accessible to a general computational science audience.

Scope (one sentence): A proof-of-concept demonstration that fragment molecular orbital calculations for helium clusters can be split across three room-temperature diamond-based two-qubit quantum computers to increase sampling throughput without losing chemical accuracy.

What This Paper Is About

Quantum chemistry on quantum hardware is limited by small qubit counts and by the sheer number of nearly identical subproblems a single calculation generates. Fragment-based methods such as FMO naturally decompose a large molecular energy into many small, independent monomer and dimer calculations — an "embarrassingly parallel" workload. The authors ask whether that structural parallelism can actually be exploited on real quantum hardware to reduce wall-clock time or raise sampling throughput, a question they state had not previously been demonstrated experimentally on physical QPUs.

Key Contributions

1. First hardware demonstration of parallel FMO quantum chemistry across multiple QPUs. Three Quantum Brilliance "Quoll" diamond-based two-qubit quantum computers were installed on-site at the Oak Ridge Leadership Computing Facility and networked to an ORNL test computing cluster, then used jointly for a chemistry workflow.

2. Two distinct parallelization protocols defined and benchmarked. A synchronous shot-parallel mode, where multiple QPUs sample the same QWFS circuits and their SPAM-corrected counts are merged into one probability distribution, and an asynchronous fragment-parallel mode, where QPUs independently evaluate different FMO fragments using device-specific optimized circuits.

3. Quantified parallel efficiency and accuracy trade-offs. The three-QPU shot-parallel workflow reaches 93.60% parallel efficiency; the fragment-parallel workflow measures 75.01%, rising to 95.17% when non-useful idle time from non-interruptible submitted jobs is excluded. In both cases assembled FMO–QWFS energies remain chemically accurate relative to the classical ideal.

4. Analysis of heterogeneous device noise on parameter transfer. The authors show that differing noise profiles across devices can degrade shot-parallel sampling when a single parameter set is transferred, and that device-specific or jointly optimized parameters restore accurate energies.

Main Findings

• Shot-parallel efficiency of 93.60%: Distributing the sampling of QWFS monomer and dimer circuits across the three Quoll QPUs and merging the corrected counts into a single probability distribution increases average shot rate close to the ideal parallel rate, with 93.60% measured parallel efficiency.

• Fragment-parallel efficiency of 75.01% (95.17% excluding idle time): Asynchronously assigning different FMO fragments to different processors yields lower measured efficiency, but 95.17% once idle time caused by non-interruptible job submission is removed from the accounting.

• Accuracy preserved relative to the classical ideal: Assembled FMO–QWFS energies for helium clusters He_n (n = 2, 4, 8, 10, 14) remain within the chemical accuracy threshold of 1.4·10⁻³ a.u. in both parallel modes. The authors stress that accuracy is measured against a deliberately restricted selected-CI reference chosen to fit the two-qubit hardware, not against full FCI.

• Heterogeneous noise affects parameter transfer: Because the three devices have measurably different SPAM confusion matrices and gate fidelities, a parameter set optimized on one device does not transfer cleanly in shot-parallel sampling; device-specific or jointly optimized parameters recover accurate energies.

• Device asymmetry is substantial: Measured SPAM-corrected single-qubit (SMSQ) average gate fidelities span 0.95–0.98 across the three systems, while native anti-controlled-Z (ACZ) fidelities span a much wider 0.79–0.96. Quoll-2's collection efficiency is almost double Quoll-1's, and Quoll-3 required higher microwave power plus added gate delays to limit diamond heating, reducing its available shot count.

• Small circuits suffice: Monomer fragments required only a one-qubit QWFS instance (a single parameterized R_y rotation) and dimers used the full two-qubit space (three R_y rotations plus the native anti-controlled controlled-Z). Each circuit execution used 5,000 shots.

Methodology in Plain English

The team picked helium clusters as a tractable testbed. For each cluster size, they generated candidate geometries, optimized them at the CAM-B3LYP/6-31+G(d) level in Gaussian 16, and kept the lowest-energy structure. GAMESS then performed FMO2 partitioning with Hartree–Fock and the cc-pVDZ basis set — for example, He₄ was split into four atomic fragments (NFRAG=4, NACUT=1) — with no electrostatic approximations or cutoffs, generating FCIDUMP integral files for every monomer and dimer.

Each fragment was then reduced to a tiny quantum problem: a restricted set of electron configurations small enough to fit on a two-qubit device. A classical selected-CI calculation on that same restricted space supplied the "ideal" reference energy.

On the quantum side, the Quantum Wave Function Sampler uses the QPU only to sample basis-state probabilities from a short parameterized circuit; the signs of the wavefunction coefficients are learned classically by a single neural-network node with a tanh activation, trained for 50 epochs at a learning rate of 0.09. The energy is then computed classically from those sampled probabilities and precomputed Hamiltonian matrix elements — meaning the quantum device never needs Pauli-basis rotation circuits.

Hardware error was handled with one method only: SPAM correction. Confusion matrices were measured with 10,000 shots per prepared basis state on each device, and corrected probability distributions were obtained by maximum-likelihood estimation under positivity and normalization constraints. Because re-optimizing the circuit separately for every fragment was impractical given device stability, the team optimized one representative monomer and one representative dimer circuit per device and held those parameters fixed, then reused the measured distributions across fragments. A fragment was accepted when its deviation from the classical ideal fell below ε_tol = 10⁻⁴ a.u.; otherwise it was resampled, for up to 50 iterations.

Why This Matters

Impact on research: The work reframes a familiar classical HPC strategy — fragmentation and massive parallelism — as a route to practical quantum chemistry. It provides the first hardware evidence that the task parallelism inherent to FMO translates into real throughput gains, and it introduces a quantitative efficiency accounting (including the distinction between raw efficiency and efficiency excluding job-scheduling idle time) that future multi-QPU studies can adopt. It also shows that QWFS-style sampling, which needs only one circuit rather than many Pauli measurement bases, is well matched to small, slow, noisy prototype devices.

Real-world applications:

• Molecular cluster and noncovalent interaction modeling, where weak interactions between many units dominate the physics.
• Biomolecular and pharmaceutical fragment screening, where drug-like systems are already routinely split into fragments for classical FMO calculations.
• Battery and catalysis materials, extending fragment-based quantum workflows toward systems such as LiCoO₂, already treated with FMO-VQE in prior work.
• HPC-integrated quantum acceleration, where QPUs act as a distributed accelerator resource inside existing supercomputing job schedulers.

Industry relevance: The demonstration is co-authored by Quantum Brilliance and Oak Ridge National Laboratory, and specifically addresses operational realities of HPC deployment — 19-inch rack mounting, 98.4 kg per unit, 339 × 436 × 1020 mm dimensions, roughly 400 W per QPU, standard 10 A mains power over three IEC-C13 cables, plus OS compliance, rack monitoring, and remote support. Room-temperature diamond operation removes the dilution refrigeration burden that dominates superconducting deployments, which matters for facilities planning to co-locate many QPUs with classical compute.

Future Directions

• Scaling to more and better QPUs: The authors motivate future HPC environments with many fault-tolerant quantum processors acting as a distributed resource, which requires both larger devices and scheduling infrastructure that avoids the idle time that cost them roughly 20 percentage points of fragment-parallel efficiency.
• Moving beyond fixed representative parameters: Because full per-fragment re-optimization was impractical, an open question is how to schedule iterative quantum optimization across many fragments without sacrificing calibration stability.
• Characterizing and mitigating heterogeneous noise more systematically: The interaction between device-specific noise and parameter transfer needs a general strategy, beyond device-specific or joint optimization, for clusters where devices differ as much as Quoll-1, Quoll-2, and Quoll-3 do.
• Extending beyond the restricted CI space: All reported accuracies are against a deliberately truncated selected-CI target that fits two qubits; recovering a larger share of the full FCI correlation energy remains untested and will require both more qubits and deeper circuits.

Target Audience

Quantum chemistry method developers and quantum algorithm researchers working on near-term and early fault-tolerant applications; HPC facility architects and operations staff evaluating on-premises QPU integration; quantum hardware vendors interested in multi-QPU workload orchestration and noise characterization; and graduate students or advanced undergraduates in computational chemistry seeking a concrete, hardware-grounded example of how fragmentation strategies map onto quantum execution.

Authors’ abstract

Fragment molecular orbital (FMO) quantum chemistry offers a natural route to parallel quantum computing: large molecular calculations can be decomposed into smaller quantum subproblems that can, in principle, run concurrently. Their practical benefit on real quantum hardware, however, has remained unclear. We demonstrate parallel FMO calculations on three room-temperature diamond-based 'Quoll' quantum computers installed on-site at the Oak Ridge Leadership Computing Facility. Helium clusters (He$_{n}$, $n=2,4,8,10,14$) served as a testbed for total energy calculations using the Quantum Wave Function Sampler (QWFS), a selected configuration interaction (SCI) approach. We compare two execution modes: synchronous shot-parallel execution, in which quantum processors jointly sample QWFS circuits and combine their SPAM-corrected counts into a single probability distribution; and asynchronous fragment-parallel execution, in which processors independently evaluate FMO fragments using device-specific optimized circuits. Both modes increase quantum sampling throughput without systematic loss of accuracy. The three-QPU shot-parallel workflow reaches a parallel efficiency of $93.60\%$, while the fragment-parallel workflow achieves a measured efficiency of $75.01\%$, increasing to $95.17\%$ when excluding the non-useful idle time associated with non-interruptible submitted jobs. In both cases, the assembled FMO-QWFS energies remain chemically accurate compared to the classical ideal. These results provide a proof-of-concept demonstration that parallel quantum execution can improve the practical throughput of quantum chemistry workflows on prototype hardware, motivating the development of future HPC environments with many fault-tolerant quantum processors operating as a distributed resource for molecular and materials simulation.