Benchmarking Quantum Architecture Search with Surrogate Assistance

TL;DR

Benchmarking Quantum Architecture Search (QAS) is computationally intensive and lacks a unified pipeline. SQuASH introduces a surrogate-based benchmark that predicts PQC performance after training, enabling fast evaluation across two core tasks—state preparation toward $|\text{GHZ}_3\rangle$ and linear classification—over two predefined search spaces. The contributions include an extendable benchmark with evaluation protocols, pretrained surrogate models (Graph Neural Network and Random Forest), and open data/code to support reproducibility. This work demonstrates substantial speedups in QAS exploration and provides a foundation for reproducible, benchmark-driven AutoQML development capable of scaling to more complex PQC architectures.

Abstract

The development of quantum algorithms and their practical applications currently relies heavily on the efficient design, compilation, and optimization of quantum circuits. In particular, parametrized quantum circuits (PQCs), which serve as the basis for variational quantum algorithms~(VQAs), demand carefully engineered architectures that balance performance with hardware constraints. Despite recent progress, identifying structural features of PQCs that enhance trainability, noise resilience, and overall algorithmic performance remains an active area of research. Addressing these challenges, quantum architecture search (QAS) aims to automate the design of problem-specific PQCs by systematically exploring circuit architectures to optimize algorithmic performance, often with varying degrees of consideration for hardware constraints. However, comparing QAS methods is challenging due to the absence of a unified benchmark evaluation pipeline, and the high resource demands. In this paper, we present SQuASH, the Surrogate Quantum Architecture Search Helper, a benchmark that leverages surrogate models to enable uniform comparison of QAS methods and considerably accelerate their evaluation. We present the methodology for creating a surrogate benchmark for QAS and demonstrate its capability to accelerate the execution and comparison of QAS methods. Additionally, we provide the code required to integrate SQuASH into custom QAS methods, enabling not only benchmarking but also the use of surrogate models for rapid prototyping. We further release the dataset used to train the surrogate models, facilitating reproducibility and further research.

Figures (11)

• Figure 1: Integration of the SQuASH benchmark and its artifacts into a custom QAS pipeline.
• Figure 2: Circuit fidelity distribution in training data for state preparation search spaces.
• Figure 3: Data collection process for ls_base. A candidate PQC is iteratively constructed by sampling gates or layers from predefined gate and layer sets. Each generated PQC is trained on the full training subset and then evaluated on the test subset. If the initial (i.e., untrained) circuit, along with its parameters, has not been encountered previously, it is stored in the dataset and saved as a persistent file for future use.
• Figure 4: Conversion of a quantum circuit with $n=3$ qubits into a DAG. Top: The original quantum circuit along with the gate set alphabet. Bottom: The corresponding DAG representation, where each node represents a quantum gate and directed edges capture temporal dependencies along qubit lines. Also shown are (i) the node feature vector for $v_7$, (ii) the edge feature vector for $e_4$ and (iii) the adjacency matrix $A_G$ for the graph.
• Figure 5: Graph-based training pipeline for the surrogate model