Quantum Computer Errors Mapped with New Learning Technique

A new method models and understands noise in quantum computers, a key barrier to building practical quantum technologies. Ryo Sakai and Yu Yamashiro present a technique for learning hardware imperfections directly from experimental data using differentiable Kraus representations and tensor networks. Their approach, tested on the ibm_fez superconducting processor, successfully reproduces observed device behaviour not only for the ripple-carry adder circuit used during training, but also generalises to a different multiplier circuit. The framework captures fundamental device characteristics and offers a pathway towards more accurate noise-aware predictions for complex quantum algorithms, such as the quantum approximate optimisation algorithm with error detection. Learned noise models predict quantum circuit behaviour across diverse architectures A fidelity exceeding 0.99 on circuits up to ten qubits was previously achievable with tensor networks, but this required complete knowledge of the quantum system and custom experimental protocols. This new method bypasses those limitations by learning a device-wide noise model directly from standard circuit execution data, a feat previously unattainable. The challenge in quantum computing lies not only in controlling qubits but also in mitigating the inevitable errors that arise from their interaction with the environment and imperfections in the hardware. These errors, collectively known as noise, corrupt quantum information and limit the coherence time, the duration for which a qubit can maintain its superposition state. Traditional methods for characterising this noise, such as quantum process tomography, are often resource-intensive, requiring many carefully calibrated measurements for each gate. Furthermore, these methods typically focus on characterising individual gate errors, neglecting the complex interplay between gates and potential crosstalk. Employing differentiable Kraus representations and tensor networks has created a framework capable of simulating a quantum device’s output distribution from a single experiment, offering a major leap in noise modelling. This is achieved by representing the noise as a set of probabilistic transformations acting on the quantum state, allowing the model to learn the underlying noise characteristics from observed data. The learned model, initially trained on a ripple-carry adder circuit on the ibm_fez processor, accurately predicted the output distribution of an entirely separate multiplier circuit without any further training, indicating its capacity to capture fundamental device characteristics. Following initial training on a ten-qubit adder circuit, the learned noise model accurately predicted the output distribution of a thirteen-qubit multiplier circuit, suggesting the framework captures inherent device characteristics rather than simply memorising the training data. This ability to generalise to different circuits is crucial for practical applications, as it allows the model to predict the performance of new algorithms without requiring extensive retraining. The method utilises differentiable Kraus operators, a mathematical tool for describing quantum noise, and tensor networks, a way of efficiently simulating quantum systems, to model noise channels attached to each gate and potential crosstalk between qubits. Tensor networks provide a compact and efficient way to represent the high-dimensional state space of a quantum system, enabling the simulation of circuits with a larger number of qubits than would be possible with traditional methods. Validation on synthetic data revealed a Choi process fidelity exceeding 0.99 when recovering ground-truth noise parameters, including complex entangling noise affecting multiple qubits. This high fidelity demonstrates the accuracy of the learned noise model in capturing the underlying noise characteristics of the quantum device. Despite representing a substantial advance in noise modelling, the current fidelity scores do not yet translate directly into guaranteed performance improvements for complex quantum algorithms requiring strong error correction. Further research is needed to integrate this noise model into error mitigation and correction schemes to realise its full potential. Differentiable Kraus operators model noise on superconducting quantum hardware Representing quantum noise using differentiable Kraus operators underpins this advance, a set of instructions describing how a quantum state is altered by noise, similar to a recipe detailing how ingredients change during cooking. These operators mathematically describe the transformation of a quantum state due to noise, allowing for a probabilistic representation of errors. Unlike deterministic transformations, Kraus operators account for the inherent randomness of quantum noise. These operators are generated via a Stinespring-based parameterization, a mathematical tool ensuring the noise model behaves realistically within the rules of quantum mechanics, guaranteeing physical plausibility throughout the learning process. The Stinespring dilation theorem provides a way to construct a completely positive and trace-preserving map, ensuring that the noise model adheres to the fundamental principles of quantum mechanics. A method was developed to model quantum hardware noise using the ibm_fez, a Heron-generation superconducting processor. Superconducting qubits, while promising, are susceptible to various noise sources, including energy relaxation, dephasing, and control errors. This approach differs from traditional quantum process tomography, which is expensive and gate-by-gate, and tensor network methods that learn the overall process rather than individual gate noise. Quantum process tomography requires a complete set of measurements to characterise the quantum channel, making it impractical for large-scale systems. Tensor network methods, while efficient, often treat the entire circuit as a single process, making it difficult to isolate and characterise individual gate errors. This new method offers a more efficient and targeted approach to noise modelling, focusing on learning the noise characteristics of individual gates and potential crosstalk between them. Predicting quantum circuit errors through data-driven transfer learning Scientists are edging closer to building quantum computers durable enough to tackle complex problems, but accurately modelling the errors within these machines remains a formidable challenge. The pursuit of fault-tolerant quantum computation requires a deep understanding of the noise characteristics of quantum hardware and the development of effective error correction strategies. This new technique, which learns noise characteristics from routine circuit data, sidesteps the need for painstaking, resource-intensive measurements of individual quantum gates, although the authors acknowledge a key limitation. Their demonstration of transfer learning, successfully predicting the behaviour of one circuit based on training with another, currently relies on relatively simple algorithms, a ripple-carry adder and a multiplier. While promising, the ability to generalise to more complex algorithms remains an open question. The ripple-carry adder and multiplier circuits, though useful for demonstrating the feasibility of the approach, represent a limited subset of the algorithms that will be required for practical quantum computation. This technique offers an important step towards understanding and modelling the inherent imperfections within quantum processors, sidestepping the need for exhaustive individual gate analysis. By learning noise characteristics from existing circuit data, scientists can more efficiently predict behaviour and assess the feasibility of more complex algorithms, such as quantum optimisation. This work presents a new method for modelling imperfections in quantum computers by learning directly from experimental data, bypassing the need for detailed analysis of individual quantum gates. Representing noise as ‘channels’, transformations affecting quantum information, and using a technique called Kraus operators allows the system to accurately simulate a device’s behaviour from a single experiment. In particular, a model trained on a ripple-carry adder successfully predicted the output of an entirely different multiplier circuit without further adjustment, indicating it captures fundamental device characteristics. The ability to accurately predict the behaviour of different circuits from a single training run represents a significant advance in noise modelling and opens up new possibilities for developing more robust and reliable quantum algorithms. Future work will focus on extending this method to more complex circuits and exploring its potential for integration with error mitigation and correction schemes. The research successfully demonstrated a method for learning the noise characteristics of a quantum processor from a single experiment. This approach models noise as ‘channels’ and uses Kraus operators to accurately simulate device behaviour, offering an alternative to analysing each quantum gate individually. Importantly, a model trained on a ripple-carry adder circuit accurately predicted the output of a multiplier circuit, suggesting it captures fundamental device properties. The authors intend to extend this method to more complex algorithms and explore its use with error mitigation techniques. 👉 More information🗞 Quantum hardware noise learning via differentiable Kraus representation on tensor networks🧠 ArXiv: https://arxiv.org/abs/2604.20804 Tags: