AI Agents for Variational Quantum Circuit Design

TL;DR

This work introduces an autonomous agent-based framework for VQC architecture search that integrates high-level reasoning with a quantum simulation environment, and demonstrates that agentic AI can effectively navigate and refine the VQC design landscape with minimal human intervention.

Abstract

Variational quantum circuits (VQCs) constitute a central building block of near-term quantum machine learning (QML), yet the principled design of expressive and trainable architectures remains a major open challenge. The VQC design space grows combinatorially with the number of qubits, layers, entanglement structures, and gate parameterizations, rendering manual circuit construction inefficient and often suboptimal. We introduce an autonomous agent-based framework for VQC architecture search that integrates high-level reasoning with a quantum simulation environment. The agent proposes candidate circuit architectures, evaluates them through fully automated training and validation pipelines, and iteratively improves its design strategy via performance-driven feedback. Empirically, we show that the agent autonomously evolves circuit architectures from simple initial ansätze toward increasingly expressive designs, progressively trying to improve task performance. This demonstrates that agentic AI can effectively navigate and refine the VQC design landscape with minimal human intervention, providing a scalable methodology for automated quantum model development in the Noisy Intermediate-Scale Quantum (NISQ) regime.

Figures (12)

• Figure 1: The main AI Agent loop for VQC design.
• Figure 2: Quantum neural network (QNN) architectures. (a) A simple QNN where the entire input is linearly embedded into $n_\text{enc}$ encoding dimensions, scaled to $[0, \pi]$, and processed by a variational quantum circuit (VQC). The VQC outputs are linearly combined and passed through a sigmoid activation to yield the final prediction. (b) A quanvolutional neural network (QuanvNN), where 1D input data is processed by sliding windows of data points and some stride, each encoded into a number of qubits and processed by a VQC with measurements. The resulting sequence is first input into a residual CNN block followed by adaptive pooling along the sequence dimension. In the residual CNN block, the measurements from the VQC are seen as channels. The output is fed into a classical neural network with two hidden layers (sizes 10 and 5, leaky ReLU activations), followed by a sigmoid output. (c) Full Quantum QNN architecture where the entire 1D input data of 21 points has to be encoded into $n_q<21$ qubits. All models are trained using mean squared error loss between predicted and true peak positions.
• Figure 3: Sample of 1D Data with Gaussian peaks: This figure illustrates examples of the 1D input data used in the quantum neural network (QNN). Each data sample contains a single Gaussian peak characterized by varying height, width, and position. The dashed, vertical line indicates the true position of the peak, which the QNN aims to predict.
• Figure 4: Claude 3.7 Sonnet Simple QNN results. The agent was allowed to run for 10, 10, 3, 1, 2 iterations with interactions between. The text interactions are documented in the text and indicated by stars in the trajectory (a) and dashed vertical lines in (b). (a) Test RMSE vs. Number of VQC Parameters: Trajectory in (RMSE, number of trainable parameters in the VQC)-space for the VQCs designed by the agent. (b) Test RMSE vs. Iteration Index: A plot of the test RMSE of the QNN after training with the VQC designed by the agent at each iteration.
• Figure 5: Quantum circuit design of the best model for the SimpleQNN found by Claude 3.7 Sonnet.