🔬 physics

Co-Design of a Quantum Circuit using QGHNN for High Separate Fidelity and Maximum Grover Algorithm Accuracy

This paper presents a hybrid physics–machine-learning framework utilizing a Quantum Graph Hamiltonian Neural Network (QGHNN) to co-design superconducting quantum circuits, achieving over 99.4% readout separation fidelity and approximately 86% Grover algorithm accuracy comparable to IBM's Q5 Tenerife processor.

Original authors: Mahdi Esmaeili, Katayoun Emadzadeh, M. Hossein Eskandari, Muhamad. B Barfar, Saeed Hajihosseini, Ahmad Salmanogli, Hesam Zandi, Abolfazl Eskandari

Published 2026-09-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahdi Esmaeili, Katayoun Emadzadeh, M. Hossein Eskandari, Muhamad. B Barfar, Saeed Hajihosseini, Ahmad Salmanogli, Hesam Zandi, Abolfazl Eskandari

Original paper licensed under CC BY 4.0 (https://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

Quantum computers promise to solve problems that would take ordinary machines thousands of years to finish, but building them is like trying to balance a house of cards in a hurricane. These machines rely on delicate units of information called qubits, which can exist in multiple states at once, allowing them to explore many possibilities simultaneously. However, this delicate state is easily shattered by noise, heat, or even the act of looking at it too closely. To make a quantum computer useful, engineers must design circuits that keep these qubits stable long enough to perform calculations, while also being able to read the results clearly before the information fades away. The challenge is that the settings that make a qubit easy to read often make it harder to perform complex math, and the settings that help it calculate well can make the final answer fuzzy. Finding the right balance requires navigating a landscape where improving one part of the system often hurts another.

In a new study, researchers at the Iranian Quantum Technologies Research Center and collaborating institutions have developed a method to solve this balancing act for a specific type of quantum computer built from superconducting circuits. They created a hybrid framework that combines the laws of physics with a specialized type of artificial intelligence to design a two-qubit system that works better than previous attempts. Instead of treating the hardware design and the algorithm performance as separate problems, the team used a neural network—a computer program designed to learn from data—to find the perfect operating point where both the measurement accuracy and the calculation success rate are maximized. Their work demonstrates that by training this system on the underlying physics of the device, they can identify a sweet spot where the machine reads its own state with high precision while also executing a famous search algorithm with remarkable success.

The researchers focused on a setup where two superconducting qubits are connected through a shared resonator, a component that acts like a bridge for their interactions. In this environment, the qubits are subject to constant noise and energy loss, which are modeled using equations that describe how open systems evolve over time. The team needed to tune several physical variables, such as the frequency of the qubits and the strength of their connection to the bridge, to achieve two goals at once: making the readout of the qubit state as clear as possible and ensuring the Grover search algorithm, a method for finding a specific item in a large list, works correctly. The difficulty lies in the fact that these two goals often pull in opposite directions; pushing the system to be more accurate in its calculations can make the readout signal blurrier, and vice versa.

To navigate this conflict, the team employed a tool called a Quantum Graph Hamiltonian Neural Network. Unlike standard computer programs that treat data as a black box, this network is built to understand the structure of the physical system, treating the qubits and their connections as a map of interacting parts. The researchers generated a large dataset by simulating the behavior of the quantum system under thousands of different physical conditions, recording how each change affected both the readout quality and the algorithm's success. They then trained the neural network to learn the complex, non-linear relationship between the physical settings and the final results. Once the network learned this map, it could predict the outcome of any new setting instantly, without needing to run the slow, heavy simulations every time. This allowed the team to use a mathematical search method to climb the hill of possibilities and find the peak where both performance metrics were high.

The results of this co-design approach were striking. Before the optimization, the system showed a clear separation between the different states of the qubit, but there was still some overlap that could lead to errors. After the neural network guided the tuning, the separation became sharper, pushing the measurement fidelity to exceed 99.4 percent. While this might seem like a small numerical jump, in the world of quantum computing, such a fraction of a percent represents a massive reduction in the chance of a wrong answer. More importantly, the algorithmic performance saw a dramatic leap. The success probability of the Grover search algorithm, which had been limited by the initial hardware settings, rose to approximately 86 percent. This level of accuracy is comparable to what is currently achieved by established commercial quantum processors, suggesting that the new design is viable for real-world applications.

The study also revealed a fundamental trade-off that engineers must manage. When the researchers optimized the system solely for the search algorithm, they found that the strong interactions needed for fast calculations tended to suppress the delicate quantum connections required for entanglement, a resource where qubits become linked in a way that defies classical physics. The neural network successfully found a middle ground, preserving enough of these quantum links while still delivering high algorithmic performance. To prove that their findings were not just theoretical, the team took the optimized parameters and used them to design a physical circuit layout in professional engineering software. They then ran electromagnetic simulations on this design, confirming that the numbers produced by the artificial intelligence translated directly into a buildable, functional hardware configuration.

This work highlights a shift in how quantum computers are built. Rather than designing the hardware first and then trying to make the software work with it, or optimizing the software and hoping the hardware can keep up, this approach designs both together. By using a machine learning model that respects the laws of physics, the researchers showed that it is possible to find operating points that satisfy multiple, competing demands simultaneously. The study confirms that even in a noisy, imperfect world, careful co-design can extract high performance from quantum circuits, paving the way for more reliable and powerful machines in the future. The ability to translate these optimized settings directly into physical dimensions, such as the length of wires or the size of capacitors, means that the path from a computer simulation to a working chip is becoming shorter and more direct.

Technical Summary: Co-Design of a Quantum Circuit using QGHNN for High Separate Fidelity and Maximum Grover Algorithm Accuracy

Problem Statement
The optimization of superconducting quantum circuits faces a critical challenge: the relationship between low-level physical device parameters (e.g., frequencies, coupling strengths) and high-level algorithmic performance is highly nonlinear and often counterintuitive. Traditional optimization strategies typically treat hardware calibration (such as maximizing readout fidelity) and algorithmic performance (such as Grover search success) as separate objectives. However, operating points that maximize measurement distinguishability do not necessarily yield high algorithmic success, and vice versa. Furthermore, existing methods often rely on brute-force simulation or heuristic tuning, which are computationally expensive and fail to capture the complex trade-offs inherent in open quantum systems subject to decoherence and dissipation.

Methodology
The authors propose a physics-informed, machine-learning-assisted co-design framework that jointly optimizes state-separation fidelity and Grover algorithm accuracy at the Hamiltonian level. The methodology integrates three core components:

Physical Modeling: The system is modeled as a two-qubit superconducting architecture (transmons) coupled via a bus resonator. One qubit is flux-tunable, while the other is fixed. The dynamics are described by a total Hamiltonian ( $H = H_{qubit} + H_{res} + H_{int}$ ) incorporating dispersive qubit-resonator interactions and effective cross-Kerr ( $ZZ$ ) couplings. Realistic open-system dynamics, including photon loss, energy relaxation, and dephasing, are simulated using the Lindblad master equation.
Performance Metrics:
- Readout Fidelity: Evaluated by simulating the dispersive measurement process to generate I/Q distributions. The separation fidelity is calculated by projecting these distributions onto an optimal discrimination axis to estimate single-shot state distinguishability.
- Algorithmic Accuracy: Assessed by simulating a two-qubit Grover search circuit. The algorithm utilizes effective $ZZ$ interactions derived from the circuit parameters to implement the oracle and diffusion operators. Success probability is measured under Lindblad dynamics.
QGHNN Surrogate and Optimization: A Quantum Graph Hamiltonian Neural Network (QGHNN) is employed as a differentiable surrogate model. The network maps Hamiltonian-level parameters (input layer: 9 neurons corresponding to the nine Hamiltonian-level parameters) to the dual performance objectives (output layer: separation fidelity and Grover accuracy).
- The QGHNN architecture is designed as a compact fully connected surrogate model consisting of an input layer with nine neurons, followed by two hidden layers with sixteen neurons each. Hyperbolic tangent (Tanh) activation functions are used in the hidden layers, and a Sigmoid activation function is used in the output layer to constrain predictions within [0, 1].
- The model is trained on a dataset of 1,200 physically constrained samples generated from Lindblad simulations.
- Once trained, the QGHNN replaces computationally intensive quantum simulations, enabling gradient-based optimization (using the Adam optimizer) directly in the parameter space to find operating points that maximize the sum of both objectives.

Key Contributions

Unified Co-Design Framework: The paper introduces a workflow that simultaneously optimizes hardware readout performance and algorithmic success, moving away from isolated optimization strategies.
QGHNN for Quantum Hardware: The application of QGHNNs provides a structured approach to modeling system parameters and performance metrics. While the network is implemented as a fully connected surrogate, the framework is grounded in Hamiltonian-level modeling to reflect the physical system's dynamics.
Trade-off Identification: The framework explicitly identifies and navigates the trade-offs between competing metrics. For instance, it demonstrates that maximizing Grover accuracy via strong $ZZ$ interactions can inadvertently suppress coherent mode hybridization and entanglement, a trade-off the co-design approach balances.
Hardware-Validated Parameters: The optimized parameters are validated through Qiskit Metal and EPR-based electromagnetic simulations. The study confirms that the QGHNN outputs are physically meaningful quantities; however, the direct translation of these optimized parameters into specific professional scaling-up design parameters (such as detailed capacitances and inductances for large-scale circuits) is identified as a direction for future work.

Results

Readout Fidelity: The QGHNN-guided optimization improved the readout separation fidelity from approximately 99.38% to 99.44%. The authors note that while this absolute increase is modest, it is significant because it was achieved without sacrificing algorithmic performance, representing a balanced operating point rather than a local readout optimum.
Grover Algorithm Accuracy: The success probability of the two-qubit Grover search increased significantly, reaching approximately 86% at the optimized design point. This performance is comparable to results achieved on the IBM Q5 Tenerife processor.
Shot Count Analysis: The study confirms that increasing the number of shots (from 256 to 8192) improves the repeatability and statistical reliability of the results but does not inherently increase the maximum achievable accuracy, which remains bounded by hardware-induced errors (gate infidelity, decoherence).
Entanglement Trade-off: Spectral analysis revealed that optimizing for high Grover accuracy (via strong $ZZ$ coupling) flattens the resonator transmission spectrum and reduces avoided crossings, indicating a degradation of steady-state entanglement. This highlights the necessity of tunable couplers for future scalable designs.
Validation: The optimized parameters were implemented in Qiskit Metal, and the resulting circuit layout was confirmed via EPR simulations, validating the physical realizability of the QGHNN outputs.

Significance
The paper claims that this work establishes a scalable, hardware-aware strategy for near-term quantum processors. By integrating open-system quantum dynamics, algorithm-level evaluation, and machine-learning-assisted optimization, the framework identifies globally favorable operating regimes that would be difficult to find through heuristic approaches. The significance lies in demonstrating that even fractional improvements in readout fidelity, when achieved concurrently with substantial gains in algorithmic performance, are practically meaningful for reducing logical error accumulation in error-corrected systems. The work concludes that the QGHNN outputs are physically meaningful quantities directly translatable into hardware design parameters, bridging the gap between theoretical optimization and practical circuit implementation.

Technical Summary: Co-Design of a Quantum Circuit using QGHNN for High Separate Fidelity and Maximum Grover Algorithm Accuracy

More like this