⚛️ PennyLane DQAS

Differentiable Quantum Architecture Search

Automated variational circuit design via end-to-end gradient-based architecture search — built on PennyLane + PyTorch

Overview • Methodology • Installation • Quick Start • Experiments • Architecture • Papers • Ecosystem

---

---

Overview

Differentiable Quantum Architecture Search (DQAS) automates the design of variational quantum circuits by treating architectural choices (which gates to apply at each position) as differentiable parameters. Instead of manually designing ansätze, DQAS discovers optimal circuit structures via gradient-based bi-level optimization.

This repository implements the core algorithm from arXiv:2505.09653 (Differentiable Quantum Architecture Search in Quantum-Enhanced Neural Network Parameter Generation, Chen et al., 2025) using PennyLane for quantum circuit simulation and PyTorch for differentiable architecture optimization.

Key features:

• Operation pool — configurable set of candidate gates (RX, RY, RZ, H, CNOT, CZ, etc.)
• Gumbel-Softmax sampling — differentiable discrete architecture selection
• Bi-level optimization — alternating inner (circuit params) and outer (arch params) loops
• Automatic architecture derivation — extract the final circuit structure after search
• Extensible design — swap in custom gate pools, layer topologies, or tasks

Why DQAS?

Challenge	DQAS Solution
Manual ansatz design requires quantum expertise	Automated discovery from an operation pool
Barren plateaus from poor circuit structure	Architecture search finds trainable circuits
Fixed circuits can't adapt to task difficulty	Task-specific architecture optimization
Discrete architecture choices non-differentiable	Gumbel-Softmax relaxation enables end-to-end gradients

---

Methodology

flowchart TB
    subgraph Search["DQAS Search Phase"]
        A[Initialize α, θ] --> B[Sample architecture via Gumbel-Softmax]
        B --> C[Inner loop: optimize θ on train set]
        C --> D{Converged?}
        D -- No --> B
        D -- Yes --> E[Outer loop: optimize α on val set]
        E --> F{Search done?}
        F -- No --> B
        F -- Yes --> G[Derive discrete architecture: argmax α]
    end
    subgraph Eval["Evaluation Phase"]
        G --> H[Reinitialize θ for derived architecture]
        H --> I[Train θ from scratch]
        I --> J[Evaluate on test set]
    end
    Search --> Eval

Loading

Bi-Level Optimization Formulation

Inner loop:    θ* = argmin_θ   ℒ_train(θ, α)
Outer loop:    α* = argmin_α   ℒ_val(θ*(α), α)

• Inner loop: Gradient descent on circuit parameters θ to minimize training loss
• Outer loop: Gradient descent on architecture logits α to minimize validation loss
• Gumbel-Softmax trick: Makes discrete sampling differentiable: z = softmax((α + g) / τ) where g ~ Gumbel(0,1)

Architecture Representation

Each circuit position maintains logits over the operation pool:

# Architecture: [n_layers, n_positions, n_ops] logits
# n_positions = n_qubits (single-qubit) + (n_qubits-1) (two-qubit)
# During search: Gumbel-Softmax samples operations differentiably
# After search:  argmax yields the discrete architecture

---

Installation

# Clone the repository
git clone https://github.com/NullLabTests/pennylane-dqas.git
cd pennylane-dqas

# Install dependencies (choose one)
pip install -r requirements.txt       # minimal
pip install -e .                       # editable install

Requirements:

• Python ≥ 3.10
• PennyLane ≥ 0.38
• PyTorch ≥ 2.0
• NumPy ≥ 1.24
• scikit-learn ≥ 1.3

---

Quick Start

from dqas.core import DQASCircuit, OperationPool
from dqas.search import DQASSearcher
from dqas.utils import generate_synthetic_data
import pennylane as qml

# 1. Setup
pool = OperationPool(["RX", "RY", "RZ", "H", "CNOT", "CZ"])
circuit = DQASCircuit(n_qubits=4, n_layers=3, op_pool=pool)
dev = qml.device("default.qubit", wires=4)

# 2. Generate data
X_train, y_train = generate_synthetic_data(n_samples=100, n_features=4)
X_val, y_val = generate_synthetic_data(n_samples=50, n_features=4)

# 3. Run architecture search
searcher = DQASSearcher(circuit, dev)
history = searcher.search(X_train, y_train, X_val, y_val, n_epochs=30)

# 4. Evaluate discovered architecture
results = searcher.evaluate(X_val, y_val)
print(f"Test loss: {results['test_loss']:.4f}")
print(f"Architecture: {results['architecture']}")

---

Experiments

Classification

Discover an optimal circuit architecture for binary classification:

python -m experiments.classification --n-qubits 4 --n-layers 3 --n-epochs 30

Time-Series Prediction

Discover an optimal circuit for forecasting sinusoidal time series:

python -m experiments.timeseries --n-qubits 4 --n-layers 3 --n-epochs 30

All results are saved as JSON to the results/ directory.

Experiment Configuration

Parameter	Default	Description
--n-qubits	4	Number of qubits
--n-layers	3	Number of circuit layers
--n-epochs	30	Outer optimization epochs
--lr-circuit	0.05	Learning rate for circuit params
--lr-arch	0.01	Learning rate for architecture params
--seed	42	Random seed
--results-dir	results	Output directory

---

Architecture

pennylane-dqas/
├── dqas/                          # Core DQAS package
│   ├── core.py                    # OperationPool, MixedLayer, DQASCircuit
│   ├── search.py                  # DQASSearcher (bi-level optimization)
│   ├── layers.py                  # Predefined circuit layer templates
│   └── utils.py                   # Data generation, printing, counting
├── experiments/                   # Run scripts
│   ├── classification.py          # Classification experiment
│   └── timeseries.py              # Time-series experiment
├── assets/                        # Visual assets
│   ├── dqas_overview.svg          # Logo / overview diagram
│   └── methodology.svg            # Methodology pipeline diagram
├── results/                       # Experiment outputs (gitignored)
├── requirements.txt
├── setup.py
└── README.md

Key Modules

Module	Description
OperationPool	Registry of available quantum gates with arity and parameter info
MixedLayer	A circuit layer with differentiable Gumbel-Softmax architecture search
DQASCircuit	Full variational circuit composed of multiple MixedLayer instances
DQASSearcher	Bi-level optimization orchestrator (inner/outer loop)

---

Papers

This implementation is primarily based on:

Differentiable Quantum Architecture Search in Quantum-Enhanced Neural Network Parameter Generation
Samuel Yen-Chi Chen, et al. (2025)
arXiv:2505.09653 [quant-ph, cs.AI, cs.LG]

Related works:

Paper	Year	Connection
Quantum Architecture Search with Meta-learning	2021	Meta-learning for QAS initialization
Quantum circuit architecture search for VQAs	2022	Foundational QAS framework (npj Quantum Information)
An Introduction to QRL	2024	Quantum RL with architecture search elements
Evaluating NNs vs VQCs	2025	Parameter efficiency of VQCs vs classical NNs
Learning to Measure QNNs	2025	Learnable observables in quantum circuits

---

Ecosystem

This repository is part of the NullLabTests research ecosystem — three companion repos exploring quantum machine learning from theory to practice:

graph LR
    H[<b>pennylane-agent-hypotheses</b><br/>Theoretical foundations] --> D[<b>pennylane-demos</b><br/>Reference implementations]
    H --> E[<b>pennylane-agent-experiments</b><br/>Jupyter notebooks + configs]
    D --> A[<b>pennylane-dqas</b><br/>★ You are here]
    E --> A
    A --> H

Loading

Repository	Description
pennylane-agent-hypotheses	5 formal hypotheses on quantum advantage in ML (H1–H5)
pennylane-demos	Python implementations of core quantum ML algorithms
pennylane-agent-experiments	Jupyter notebooks, YAML configs, and experiment lifecycle
pennylane-dqas (this repo)	Differentiable Quantum Architecture Search — automated circuit design

---

Citation

If you use this code in your research, please cite:

@software{pennylane-dqas2025,
  author = {NullLabTests},
  title = {PennyLane DQAS: Differentiable Quantum Architecture Search},
  year = {2026},
  url = {https://github.com/NullLabTests/pennylane-dqas}
}

@article{chen2025differentiable,
  title = {Differentiable Quantum Architecture Search in Quantum-Enhanced
           Neural Network Parameter Generation},
  author = {Chen, Samuel Yen-Chi and others},
  journal = {arXiv preprint arXiv:2505.09653},
  year = {2025}
}

---

_{Built with ⚛️ by NullLabTests using PennyLane and PyTorch}