E-DQAS on PennyLane - A Demonstration of Platform-Agnostic and Hardware-Tailored Architecture Search

Abstract

The rapidly evolving quantum hardware landscape presents a strategic challenge: a commitment to any single software or hardware stack carries significant risk. To ensure long-term viability and agility, algorithmic discovery platforms must be portable. This paper demonstrates the successful migration of the Apoth3osis Enhanced Differentiable Quantum Architecture Search (E-DQAS) framework to the Pennylane quantum machine learning library. This strategic initiative validates our methodology's backend-agnosticism and begins the process of tailoring our discovery engine for specific next-generation hardware, including photonic quantum processors.

View Related Publications

GitHub Repo : https://github.com/Apoth3osis-ai/e-dqas_xanadu

Research Gate: https://www.researchgate.net/publication/392788035_E-DQAS_on_Pennylane_A_Demonstration_of_Platform-Agnostic_and_Hardware-Tailored_Architecture_Search

We deploy the re-engineered E-DQAS platform to find the ground state of the 1D Transverse Field Ising Model (TFIM), a canonical benchmark for quantum phase transitions. This study deliberately omits the advanced symmetry-enforcement penalties of our previous work to establish a baseline of the framework's raw discovery power on the new backend. The architecture search is driven by a differentiable, generator-based effective Hamiltonian, a mechanism adapted for optimal performance within Pennylane's differentiation engine.

The results show that the platform successfully discovers compact, performant circuits that accurately map the TFIM's phase diagram. This demonstrates the intrinsic power of the core E-DQAS concept and its adaptability to different quantum programming paradigms. This work positions our proprietary methodology as a portable, hardware-aware asset, ready for partnership and deployment across the quantum ecosystem to discover fundamentally new and efficient algorithmic structures.

1 Introduction

The quantum computing industry is in a state of dynamic flux, with multiple competing hardware modalities—superconducting circuits, trapped ions, photonics, and neutral atoms—vying for long-term viability. In this environment, an algorithm development strategy tethered to a single platform is untenable. The key to future success lies in creating portable intellectual property that can be deployed on the best available hardware, regardless of the vendor or underlying physical architecture.

In line with this strategy, Apoth3osis has undertaken a successful port of our proprietary E-DQAS framework from a Qiskit/PyTorch stack to Xanadu's Pennylane library. This effort was driven by two strategic objectives:

De-risking and Portability: To prove that our core E-DQAS methodology is a flexible, backend-agnostic asset, ensuring its longevity and applicability across the quantum ecosystem.

Hardware-Specific Tailoring: To begin adapting the framework for emerging hardware platforms, with this implementation specifically designed with the gate sets and differentiation methods suitable for photonic quantum computers in mind.

This paper details the re-engineered framework and validates its performance on the 1D Transverse Field Ising Model (TFIM). The TFIM, with its well-understood quantum phase transition, serves as a perfect benchmark to test the platform's ability to find accurate ground state circuits in different physical regimes. By demonstrating success on a new backend, we showcase the power and adaptability of our core approach to automated algorithm discovery.

2 The E-DQAS Methodology on Pennylane

While the high-level concept of E-DQAS remains the same—a "guided discovery" process driven by a classical neural network—the implementation on Pennylane required adapting our differentiable mechanism to the new backend.

2.1. The Portable Classical Controller

The "brain" of E-DQAS is a set of classical neural networks, implemented in PyTorch, that remains largely unchanged across backends. It consists of:

A Gate Selector: A policy network that outputs logits for selecting a gate at each possible location in a circuit layer.

A Gate Parameter Generator: A network that generates the continuous parameters (e.g., rotation angles) for the selected gates.

This modular design ensures that the core learning component of our IP is portable.

2.2. Generator-Based Differentiable Interface

The primary adaptation for Pennylane was in the mechanism for differentiable gate application. While our previous work used a weighted sum of unitary matrices, this implementation uses a weighted sum of gate generators.

Mechanism: In quantum mechanics, any unitary gate U(θ) can be expressed as U(θ) = exp(-i H_g θ), where H_g is the gate's generator (a Hermitian operator). For each layer, we construct an effective Hamiltonian, H_layer, as a differentiable, weighted sum of the generators of all possible gates: H_layer = Σᵢ pᵢ H_gateᵢ(θᵢ) Here, pᵢ and θᵢ are the probabilities and parameters output by the classical controller. Pennylane's @qml.qnode can then simulate the evolution exp(-i H_layer t) and analytically compute gradients with respect to pᵢ and θᵢ, allowing us to train the controller. This adaptation was a practical necessity for optimal integration with Pennylane's execution model and demonstrates our ability to tailor our methods to the native capabilities of a given backend.

2.3. A Note on Symmetry

Our previous work on the TFIM introduced a sophisticated adaptive penalty to enforce the model's Z₂ symmetry. For this initial Pennylane implementation, we have deliberately omitted this constraint. The purpose is to first establish a baseline, testing the raw discovery power of the core E-DQAS framework on the new platform without the aid of physics-informed penalties. The successful re-integration of symmetry-preserving techniques is a planned next step in the development roadmap for this version of the platform.

3 Experimental Validation: TFIM Phase Diagram

We deployed the Pennylane-based E-DQAS framework to find the ground state of a 4-qubit TFIM. The experiment involved training a new model for 20 different values of the transverse field h to map out the system's behavior across its quantum phase transition.

3.1. Implementation Details

Hamiltonian: 4-qubit TFIM with J=1.0 and h scanned from 0.1 to 2.0.

Hardware Target: The simulation was run on Pennylane's high-performance lightning.gpu device.

Gate Set: The search space included gates like RX, RY, RZ, RXX, and RZZ, which have efficient decompositions on various hardware platforms, including photonic devices.

Training: Each model was trained for 100 epochs using the Adam optimizer.

3.2. Results and Analysis

The platform successfully discovered circuits that accurately tracked the ground state energy across the entire phase diagram, including the critical point at h=1.0.

The key performance indicators were all met:

Accuracy: The discovered circuits consistently found the ground state energy with high fidelity when compared to exact diagonalization results for each value of h.

Generalization: The framework demonstrated that the E-DQAS search methodology is robust enough to find effective circuits in both the ordered (h < 1) and disordered (h > 1) phases of the model, without requiring manual changes to the ansatz structure.

Efficiency: The training process for each point in the phase diagram was computationally efficient, validating the generator-based differentiable approach on the Pennylane backend.

This successful validation proves that the core principles of E-DQAS are sound and highly portable across different quantum software and hardware paradigms.

4 Potential Applications (Photonic Focus)

The successful port to Pennylane, a framework with strong ties to photonic quantum computing, opens several strategic application areas:

Photonic VQE: E-DQAS can now be used to design bespoke ansätze for solving chemistry and materials science problems on continuous-variable (CV) or specific discrete-variable (bosonic) photonic processors.

Quantum Communication: The framework can be adapted to optimize circuits for state preparation and entanglement distribution in quantum networks, where modeling and mitigating photon loss (a non-Hermitian effect) is critical.

Photonic Machine Learning: Leveraging Pennylane's strengths in QML, E-DQAS can be used to autonomously design the architecture of quantum neural networks for photonic devices, a key step in developing new AI hardware.

5 Strategic Implications and Future Work

This work is more than a technical exercise; it is a statement of strategic intent. By demonstrating the portability of the E-DQAS platform, we have validated our core IP as a flexible, backend-agnostic asset. In an industry where the winning hardware platform is still unknown, this agility is a crucial competitive advantage. We are not betting on a single horse in the quantum race; we are building the engine that any jockey can use to win.

Our future work will proceed along three parallel tracks:

Symmetry Re-integration: Implement our advanced, symmetry-preserving techniques within the Pennylane framework to further enhance the accuracy and physical validity of the discovered circuits.

Hardware Benchmarking: Engage with a hardware partner to deploy and benchmark the circuits discovered by this framework on a real photonic quantum processor.

Multi-Objective Optimization: Integrate our Pareto optimization modules to generate portfolios of hardware-specific circuits that balance accuracy against photon loss, gate errors, and other device-specific metrics.

6 Conclusion: A Call for Partnership

Apoth3osis has developed a powerful, portable, and proven engine for the automated discovery of quantum algorithms. We have demonstrated its successful deployment on two of the world's leading quantum software platforms, tailoring its internal mechanisms to best leverage the strengths of each.

This capability is a strategic asset. Our E-DQAS platform can significantly reduce the RD time and expertise required to create high-performance, proprietary quantum algorithms. We are now seeking a hardware partner, particularly in the photonic computing space, to unlock the next stage of this technology. A partnership with Apoth3osis provides an opportunity to leverage our discovery engine to build a library of circuits tailored specifically to your device, creating a powerful, defensible moat and accelerating your path to demonstrating quantum advantage on real-world problems.