HQC-Orch: Enhancing Hybrid Quantum-Classical ML through Adaptive Orchestration

Abstract

Quantum-classical Hybrid Machine Learning (ML) workflow suffer from orchestration challenges owing to static QPU resource management, resulting in inefficient resource usage and unpredictable latency. We propose HQC-Orch, an intermediary software stack that dynamically orchestrates ML computation among QPUs and classical CPUs based on the depth of quantum circuits, noise robustness, and latency requirements. HQC-Orch features a task classifier aware of noise, a quantum job scheduler that prioritizes jobs and adaptively assigns shots, and a result aggregator equipped with readout error correction. Experiments on IBM Qiskit Aer and PennyLane simulators using variational quantum eigensolver (VQE), quantum approximate optimization algorithm (QAOA), and quantum neural network (QNN) applications indicate that HQC-Orch achieves up to 57% latency reduction versus pure simulation for shallow circuits ($\leq$ 12 depths) when compared to pure simulation approaches and lower QPU occupancy rates from 100% to 30% as quantum circuit depths increase. Our approach achieves 85% QNN classification accuracy, signifying 10% sacrifice because of shot noise.