[Paper] Investigating Quantum Circuit Designs Using Neuro-Evolution

Source: arXiv - 2602.03840v1

Overview

The paper introduces EXAQC (Evolutionary eXploration of Augmenting Quantum Circuits), a neuro‑evolution framework that automatically designs and trains parameterized quantum circuits (PQCs). By treating circuit topology, gate selection, qubit connectivity, and depth as evolvable traits, the authors show that evolutionary search can produce high‑performing, hardware‑aware quantum models without hand‑crafted templates.

Key Contributions

• Unified evolutionary search over gate types, qubit connectivity, parameterization, and circuit depth, respecting device‑specific noise and topology constraints.
• Hardware‑agnostic implementation that works with both Qiskit and Pennylane, letting developers plug in any back‑end (IBM, Rigetti, etc.).
• Joint architecture‑and‑parameter optimization, eliminating the need for a separate training phase after circuit synthesis.
• Empirical validation on several classification benchmarks, achieving >90 % accuracy with modest computational budgets.
• State‑fidelity experiments demonstrating that evolved circuits can closely emulate target quantum states (high fidelity scores).

Methodology

1. Population Encoding – Each individual in the evolutionary population encodes a complete PQC: a list of layers, the gate type per layer (e.g., Rx, Rz, CNOT), the qubits each gate acts on, and the continuous rotation parameters.
2. Genetic Operators –
   • Mutation randomly changes gate types, rewires connections, tweaks rotation angles, or inserts/deletes layers.
   • Crossover swaps sub‑circuits between two parents, preserving functional blocks.
3. Fitness Evaluation – For a given task (e.g., binary classification), the circuit is executed on a simulator or real device; fitness combines task performance (accuracy or fidelity) with a penalty for depth and gate count to encourage hardware‑friendly designs.
4. Evolution Loop – The population undergoes selection (e.g., tournament), reproduction, and replacement for a fixed number of generations or until convergence.
5. Framework Integration – EXAQC wraps the Qiskit and Pennylane APIs, automatically translating the evolved genome into executable quantum programs and handling device‑specific constraints (connectivity maps, noise models).

The process is analogous to neuro‑evolution for classical neural nets, but the “neurons” are quantum gates and the “architecture” is the circuit topology.

Results & Findings

• Rapid convergence: Most runs reached >90 % accuracy within 30 generations (~200 circuit evaluations).
• Hardware awareness: When constrained to IBM’s 5‑qubit device topology, evolved circuits automatically avoided unsupported CNOTs, resulting in lower error rates than unconstrained designs.
• Parameter efficiency: Evolved circuits typically used fewer parameters than comparable manually designed ansätze, reducing training time on noisy intermediate‑scale quantum (NISQ) hardware.

Practical Implications

• Accelerated prototyping – Developers can feed a dataset and let EXAQC generate a ready‑to‑run PQC, cutting weeks of manual ansatz engineering.
• Hardware‑specific optimization – By embedding device topology and noise models, the method yields circuits that are more likely to succeed on real quantum processors, lowering the cost of cloud‑based quantum experiments.
• Cross‑platform portability – Because the framework supports both Qiskit and Pennylane, teams can switch back‑ends (IBM, Rigetti, Xanadu) without rewriting the evolutionary logic.
• Potential for AutoML‑style pipelines – EXAQC could be integrated into larger quantum AutoML systems, automatically selecting the best quantum‑classical hybrid architecture for a given problem.
• Educational tool – The evolutionary visualizations help newcomers understand how gate placement and connectivity affect performance, serving as a teaching aid for quantum circuit design.

Limitations & Future Work

• Scalability – Experiments were limited to ≤5 qubits; scaling the evolutionary search to larger registers will require smarter sampling or hierarchical evolution.
• Evaluation cost – Fitness evaluation still relies on running many circuit instances, which can be expensive on real hardware; surrogate models or transfer learning could mitigate this.
• Benchmark breadth – The study focused on classification and state‑preparation tasks; applying EXAQC to variational quantum eigensolvers (VQE) or quantum chemistry remains an open avenue.
• Hybrid search strategies – Combining gradient‑based parameter tuning with evolution (e.g., Lamarckian evolution) could further improve convergence speed.

Overall, the paper demonstrates that neuro‑evolution is a viable, hardware‑aware strategy for automating quantum circuit design, opening the door to more accessible and performant quantum machine‑learning pipelines.

Authors

• Devroop Kar
• Daniel Krutz
• Travis Desell

Paper Information

• arXiv ID: 2602.03840v1
• Categories: cs.NE, cs.LG
• Published: February 3, 2026
• PDF: Download PDF