🤖 machine learning

Investigating Quantum Circuit Designs Using Neuro-Evolution

This paper introduces EXAQC, a neuro-evolutionary framework that automates the design and training of parameterized quantum circuits by jointly optimizing gate types, connectivity, and parameters under hardware constraints, achieving high accuracy and fidelity on benchmark tasks.

Original authors: Devroop Kar, Daniel Krutz, Travis Desell

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Devroop Kar, Daniel Krutz, Travis Desell

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a complex machine out of Lego bricks to solve a specific puzzle. Usually, engineers have to manually design the blueprint, deciding exactly where every brick goes, which colors to use, and how they connect. If they get it wrong, the machine might be too big, too slow, or simply not work at all.

This paper introduces a new way to design these "machines," but instead of Lego bricks, they are quantum circuits (the blueprints for quantum computers), and instead of a human engineer, they use a digital evolution process to do the designing.

Here is the breakdown of their approach, called EXAQC, using simple analogies:

1. The Problem: The "Blind Architect"

Designing quantum circuits is incredibly hard. The number of ways you can arrange the parts (gates) and connect them (wires) is astronomical.

The Old Way: Humans try to guess the best design or use rigid templates. It's like trying to build a house by only using pre-made walls and never changing the floor plan. It often doesn't fit the specific problem or the hardware it runs on.
The New Way (EXAQC): Instead of a human architect, the authors use a "digital natural selection" process. They let the computer "evolve" the best circuit design over time, just like nature evolves animals to survive in their environment.

2. How the Evolution Works: The "Digital Darwinism"

The researchers created a system where quantum circuits are treated like living organisms with DNA (called a genome).

The DNA: The "genome" is a list of instructions telling the quantum computer what gates to use, which qubits (quantum bits) to connect, and in what order.
Mutation (Random Changes): Just like biological mutations, the system randomly changes the circuits. It might:
- Add a new gate (like adding a new room to a house).
- Disable a gate (removing a useless room).
- Swap wires (rearranging the plumbing).
- Reorder steps (changing the sequence of instructions).
Crossover (Breeding): The system takes two successful circuits and "breeds" them. It mixes their DNA to create a child circuit that hopefully inherits the best features of both parents. They even invented special ways to mix DNA from four parents at once, not just two.

3. The Training: "Learning by Doing"

Once a new circuit is born, it has to prove its worth. The system tests it in two main ways:

The "Teacher" Test (Imitation): The computer is given a "perfect" circuit (the teacher) and asked to copy its behavior. If the new circuit acts exactly like the teacher, it gets a high score. This is like a student trying to perfectly mimic a master pianist's performance.
The "Quiz" Test (Classification): The circuit is given a standard dataset (like sorting pictures of flowers or wines). It has to look at the data and make a guess. If it guesses correctly, it gets a high score. This is like a student taking a multiple-choice test.

The circuits that score the highest get to "reproduce" and pass their genes to the next generation. The ones that fail are discarded. Over hundreds of generations, the population of circuits gets smarter and more efficient.

4. The Results: What Did They Find?

The authors ran this evolutionary process on several real-world datasets (like the famous Iris flower dataset and Breast Cancer data) and on synthetic quantum puzzles.

High Scores: The evolved circuits were able to solve classification problems with over 90% accuracy on most tests.
Perfect Imitation: When asked to copy a "teacher" circuit, the evolved circuits achieved 98% to 100% fidelity (meaning they were almost identical to the target).
Surprising Designs: The circuits didn't just copy the teacher; sometimes they found different ways to solve the same problem. For example, to act like a "do nothing" (identity) gate, one evolved circuit used a specific gate twice in a row to cancel itself out, while another used a different trick entirely. This shows the system is creative, not just a copycat.

5. Why This Matters

The paper argues that this method is powerful because:

It's Flexible: It works with different software tools (Qiskit and Pennylane), so it's not tied to one specific brand of quantum computer.
It's Automated: It doesn't need a human to decide the structure; it figures out the best shape, depth, and connections on its own.
It's Efficient: It found good solutions with a relatively small number of attempts (only 500 "generations" of circuits were tested in their experiments).

In short: The paper presents a "survival of the fittest" simulator that automatically designs quantum circuits. Instead of humans guessing the best layout, the computer tries thousands of random variations, keeps the ones that work best, and breeds them until it creates a highly efficient, problem-solving quantum machine.

Technical Summary: Investigating Quantum Circuit Designs Using Neuro-Evolution

Problem Statement

The design of effective Parameterized Quantum Circuits (PQCs) remains a significant bottleneck in Quantum Machine Learning (QML). The structure of a quantum circuit—specifically its gate types, qubit connectivity, parameterization, and depth—fundamentally dictates its expressivity, trainability, and robustness to noise. Current approaches rely heavily on manually crafted templates or fixed heuristics, which often fail to generalize across different tasks or hardware constraints. These static designs can suffer from issues such as barren plateaus, weak gradient signals, or insufficient entanglement. While automated Quantum Neural Architecture Search (QNAS) exists, many existing methods are limited by shallow search spaces, high computational overhead, or a lack of flexibility in supporting different quantum backends and objective functions.

Methodology: EXAQC

The authors propose Evolutionary eXploration of Augmenting Quantum Circuits (EXAQC), a neuro-evolutionary framework for the automated design and training of quantum circuits. Inspired by the EXAMM and EXA-GP algorithms, EXAQC treats quantum circuits as mutable "genomes" that evolve to optimize both structure and parameters simultaneously.

1. Genome Representation

Unlike fixed-size representations, EXAQC utilizes an unbounded list of gates. A genome consists of:

Input and output qubit identifiers.
A list of gates, where each gate includes a depth value ( $0.0 \le d \le 1.0$ ), input/output qubit connections, trainable parameters (if applicable), and an innovation number.
The innovation number allows the system to track homologous structures across different genomes, facilitating effective crossover.

2. Evolutionary Operators

EXAQC employs a hybrid evolutionary-variational approach where structural changes are driven by evolution, while parameters are optimized via gradient-based learning (Adam optimizer).

Mutation Operators:
- Add Gate: Inserts a new gate at a random depth, selecting input/output qubits based on connectivity analysis to ensure the gate contributes to the computation.
- Enable/Disable Gate: Toggles the activity of existing gates.
- Reorder Gate: Moves a gate to a new depth.
- Swap Qubits: Reassigns a gate's qubit connections.
- Constraint Check: After mutation, a forward traversal ensures at least one input flows to an output; invalid genomes are discarded.
Crossover Operators:
- Binary Crossover: Combines two parents, preserving shared innovations and probabilistically retaining unique innovations from the fitter parent ( $best\_keep\_rate$ ) or the less fit parent ( $other\_keep\_rate$ ).
- N-ary Crossover: Extends crossover to multiple parents (e.g., $n=4$ ), using a randomized simplex approach to recombine parameters.
- Exponential Crossover: Splits parents at a random depth crossover point, taking the lower-depth gates from one parent and higher-depth gates from another.
Lamarckian Weight Inheritance: To reduce backpropagation costs, weights from successful parents are inherited and refined rather than re-initialized.

3. Backend Agnosticism and Loss Functions

The framework is designed to be backend-agnostic, supporting both Qiskit and Pennylane. It supports two primary optimization modes:

Teacher Imitation (Fidelity-Based): When a target "teacher" circuit exists, the system minimizes the loss between the candidate's state and the teacher's state. Metrics include Fidelity ( $1 - |\langle \phi | \psi \rangle|^2$ ) and Angular Distance.
Supervised Learning (Readout-Based): For classical datasets, the system uses measurement-driven losses. This includes Cross-Entropy on marginal readout probabilities for classification, Kullback-Leibler (KL) Divergence, and Observable Mean Squared Error (MSE) for hardware-aligned constraints.

Key Contributions

Unified Framework: Introduction of EXAQC, a system that jointly optimizes circuit topology (gate types, connectivity, depth) and parameters for both supervised learning and circuit emulation.
Backend Agnosticism: A flexible training pipeline compatible with both PennyLane and Qiskit, allowing the same evolutionary representation to be evaluated under different execution models.
Flexible Loss Formulations: The ability to switch between fidelity-based supervision (for target circuit emulation) and readout-based losses (for classical classification) within a consistent pipeline.
Novel Crossover Strategies: Adaptation of neuroevolutionary crossover techniques (N-ary and Exponential) specifically for ordered quantum gate lists.

Experimental Results

The authors evaluated EXAQC on four UCI benchmark datasets (Iris, Wine, Seeds, Breast Cancer) and a suite of synthetic "teacher" quantum circuits.

Classification Performance: Using a modest computational budget (500 evaluated genomes), evolved circuits achieved high accuracy on benchmark datasets. For example, the Iris dataset reached up to 90.0% accuracy, and Seeds reached 95.2%. The results indicated a correlation between circuit complexity (number of gates) and performance, suggesting that longer search times could yield further improvements.
Circuit Emulation: In teacher-based benchmarks, the system successfully learned to emulate target behaviors:
- Baseline (Single Gate): Achieved 100% fidelity within 3–5 genomes.
- Bell-State Generators: Achieved 98.32% fidelity.
- Input-Controlled Circuits: Achieved 94.11% fidelity.
- Multi-Layered Circuits: Achieved 91.73% fidelity after exploring over 800 genomes.
Structural Discovery: The system discovered non-trivial circuit topologies organically. In some cases, the evolved circuits achieved high fidelity using structurally different mechanisms than the teacher (e.g., applying a Hadamard gate twice to simulate an Identity gate).

Significance and Claims

The paper positions EXAQC as a practical and scalable methodology for discovering "problem-aware" quantum circuits. The authors claim that evolutionary search is a critical tool for advancing QML because it can navigate the combinatorial design space of gate types and connectivity without relying on fixed templates.

The significance of the work lies in its demonstration that:

Evolutionary algorithms can effectively discover circuit structures that generalize across different datasets and hardware constraints.
A unified framework can handle both physics-driven tasks (emulating quantum states) and data-driven tasks (classifying classical data) without changing the underlying evolutionary machinery.
The approach provides a principled pathway toward scalable quantum circuit design, moving beyond manual heuristic design toward automated, adaptive discovery.

The authors remain modest regarding future implications, noting that while the current implementation uses a single steady-state population and single-objective optimization, the framework is designed to be extensible to multi-objective strategies and island models in future work. They emphasize that the current results serve as a proof-of-concept for the viability of neuro-evolution in quantum architecture search.