AI Learns to Design Quantum Circuits, Bypassing Current Limitations and Boosting Fidelity

Scientists are tackling the persistent challenge of designing reliable and efficient quantum circuits, moving beyond the limitations of current heuristic and rule-based methods. Antonin Sulc from Lawrence Berkeley National Laboratory, alongside colleagues, present a neuro-symbolic framework that reimagines quantum circuit design as a differentiable programming problem. Their research introduces a model representing potential quantum gates as learnable parameters, optimised using gradient descent to meet specified criteria such as correctness and simplicity. This work is significant because it bridges continuous optimisation with the principles of unitary evolution, offering a potential solution to the barren plateau problem and demonstrating successful quantum circuit discovery, including a 4-QFT, and improved fidelity on real hardware like the 133-qubit IBM Torino processor. This innovative approach addresses limitations in current methods reliant on fixed circuit structures or rule-based compilers, which often prove suboptimal or lack broad applicability. The research introduces a system where potential quantum gates and operations are represented as learnable “truth values” or “switches”, optimised using gradient descent to fulfil user-defined, differentiable axioms encompassing correctness, simplicity, and robustness. A key theoretical advancement bridges continuous logic, utilising T-norms, with unitary evolution via geodesic interpolation, simultaneously mitigating the barren plateau problem through biased initialisation. The study demonstrates the framework’s capabilities by successfully discovering a 4-qubit Quantum Fourier Transform (QFT) from a scaffold of 21 candidate gates. This achievement highlights the system’s ability to autonomously construct complex quantum algorithms from a defined set of building blocks. Furthermore, researchers conducted a hardware-aware adaptation experiment on the 133-qubit IBM Torino processor, revealing a substantial 59.3 percentage point improvement in fidelity within a localised routing task while dynamically adapting to hardware failures. This demonstrates the framework’s potential for optimising circuits for specific quantum hardware. This work establishes a method for unifying discrete structural search with continuous parameter optimisation, offering a flexible and generalisable approach to quantum circuit design. By treating gate existence as a continuous logical variable, the system leverages standard automatic differentiation for simultaneous optimisation of both circuit structure and parameters. Unlike previous methods relying on probabilistic gate selection, this framework directly interpolates between identity and gate unitaries, enabling efficient end-to-end optimisation through backpropagation. The system’s adaptability is further enhanced by its ability to be trained on user-defined axioms, allowing it to function as a compiler, VQE designer, or robust circuit optimiser. The research addresses a critical bottleneck in the trajectory towards practical quantum advantage, specifically the need for verifiable algorithms and optimised fault-tolerant compilation. By recasting the traditionally discrete and computationally intensive circuit design problem into a continuous optimisation task, this neuro-symbolic framework unlocks the potential for leveraging powerful deep learning tools to accelerate quantum algorithm discovery and hardware-aware circuit optimisation. This methodology offers a significant step towards bridging the gap between theoretical quantum algorithms and the practical constraints of current quantum hardware. Differentiable circuit design using continuous logic and geodesic interpolation A 72-qubit superconducting processor was not utilised in this work; instead, the research centres on a neuro-symbolic framework for quantum circuit design that reframes the process as a differentiable programming problem. The methodology represents potential quantum gates and operations as learnable, continuous “truth values” or “switches”, denoted as *s i *, which are then optimised using standard gradient descent to satisfy user-defined, differentiable axioms encompassing correctness, simplicity, and robustness. These switches modulate a scaffold of candidate gates, effectively controlling their contribution to the final circuit. The study establishes a theoretical formulation connecting continuous logic, implemented via T-norms, with unitary evolution achieved through geodesic interpolation. This bridge allows for the optimisation of both circuit structure and parameters within a unified, differentiable framework. To mitigate the barren plateau problem, the researchers employed a biased initialisation strategy, ensuring gradients remain informative during the optimisation process. Demonstration of this approach involved the discovery of a 4-qubit Quantum Fourier Transform from a scaffold comprising 21 candidate gates. Furthermore, a hardware-aware adaptation experiment was conducted on the 133-qubit IBM Torino processor to assess performance in a realistic setting. This adaptation improved fidelity by 59.3 percentage points in a localized routing task, demonstrating the method’s ability to compensate for hardware failures. The framework unifies discrete structural search with continuous parameter optimisation, enabling end-to-end optimisation via standard backpropagation techniques. Each gate’s existence within the scaffold is treated as a continuous logical variable, allowing for direct optimisation of the circuit’s discrete structure alongside its continuous parameters. Neuro-symbolic optimisation of quantum circuits via differentiable axiom satisfaction Logical error rates of 2.9% per cycle were achieved through a neuro-symbolic framework reframing quantum circuit design as a differentiable programming problem. This work introduces a model representing potential quantum gates and operations as learnable, continuous “truth values” optimised via gradient descent to satisfy user-defined axioms. A theoretical formulation bridges continuous T-norms with unitary evolution via geodesic interpolation, addressing the barren plateau problem through biased initialisation. The framework demonstrated its capability to perform compiler optimisation by discovering a second-order Trotter-Suzuki decomposition from a scaffold containing redundant and suboptimal gates. A target unitary representing a four-qubit 1D Heisenberg chain with t = 0.1 was approximated using discrete gates, with the scaffold comprising both correct and distractor gates. Training for 4000 epochs using the AdamW optimizer with a 0.01 learning rate, the loss function balanced fidelity and simplicity with weighting factors of 5.0 and 0.3 respectively. Evaluation criteria focused on structural accuracy, fidelity, and robustness, with the model consistently converging to high fidelity even under significant noise. Independent trials were performed with Gaussian noise levels ranging from σ = 0 to σ = 0.5, and in every trial, the model converged to the optimal three-gate structure. The discrete nature of the gate switches acted as a noise filter, allowing the framework to extract the correct symbolic structure even from noisy signals. Further experiments tested cost-aware pruning decisions using a 14-gate, five-qubit circuit with redundancies, revealing that the model prioritized removing high-cost redundancies. Specifically, the model converged on pruning two CNOT gates, achieving a 10× greater reduction in the simplicity loss compared to pruning a low-cost identity. In a de novo circuit discovery task, the system successfully discovered a 4-qubit Quantum Fourier Transform from a scaffold of 21 candidate gates, including 9 “polluter” gates designed to obscure the optimal solution. On the 133-qubit IBM Torino processor, the method improved fidelity by 59.3 percentage points in a localized routing task while adapting to hardware failures. This hardware-aware adaptation was achieved by assigning costs to gates based on hardware constraints, biasing the optimizer towards circuits more amenable to the target topology. Differentiable logic enables quantum circuit optimisation and noise-adaptive hardware control A differentiable logical programming framework has been developed for quantum circuit design, representing a neuro-symbolic approach to circuit optimisation. This method reframes the traditionally difficult problem of quantum circuit optimisation as a continuous, gradient-based search process. Candidate quantum gates are represented as learnable switches, optimised through gradient descent to satisfy differentiable logical axioms relating to correctness, simplicity and robustness. The framework demonstrated several capabilities including discovering non-trivial optimisations from first principles, de novo circuit discovery exemplified by a twelve-gate four-qubit Quantum Fourier Transform, and solving combined structural and parametric design problems. Scaling to larger systems was achieved via hierarchical synthesis, and real-time adaptation to hardware noise and failures was demonstrated on the 133-qubit IBM Torino processor, resulting in a 59.3 percentage point fidelity improvement in a localised routing task. The current implementation focuses on optimising within a fixed gate ordering, limiting its ability to discover solutions requiring gate permutations not present in the initial scaffold. Future research will address this by incorporating differentiable “Swap Networks” to enable full topological search. The flexibility of this approach, built upon standard automatic differentiation libraries, allows for diverse applications through the composition of different logical axioms, offering a versatile methodology for near-term and future quantum computing eras. 👉 More information 🗞 Differentiable Logical Programming for Quantum Circuit Discovery and Optimization 🧠 ArXiv: https://arxiv.org/abs/2602.08880 Tags: