Quantum computing: Using AI for Calibration

Using artificial intelligence to improve and automate the calibration of quantum computers: this is the goal of a collaboration between LRZ and NVIDIA. Initial results are now available.

Quantum computers are still extremely sensitive and often produce different results for the same calculations. To improve the reliability of quantum computations’ fidelity, the devices must be calibrated regularly. At Euro-Q-Exa, this usually happens daily and at night. However, with the help of artificial intelligence (AI), this process could not only be monitored but also automated. To explore this possibility, the Leibniz Supercomputing Centre (LRZ) and NVIDIA have teamed-up for a joint research effort. The technology provider has developed an AI model called Ising specifically for evaluating calibration results and its optimisation, which the LRZ has been extensively tested in the daily operation of Euro-Q-Exa. The first experimental results will be presented in mid-September in Toronto during the IEEE International Quantum Week. We spoke with LRZ quantum specialists Dr. Xiaolong Deng and Hossam Ahmed about their experiences with Ising and their ideas for providing better support to Euro-Q-Exa users with the help of AI.

Why do quantum systems such as Euro-Q-Exa need to be calibrated regularly?
Hossam Ahmed: Every technology requires calibration. Cars and machines also need to be adjusted from time to time. Quantum computing, however, is a novel technology that requires more frequent calibration because the metrics describing the system can change rapidly during operation. As a result, calculations performed on a quantum computer do not always yield the same result. Changes may develop over minutes, hours, or throughout the day, which is why important parameters must be reassessed continuously. Together with IQM Quantum Computers, for example, we estimate how quickly Euro-Q-Exa’s operating frequencies change and perform calibration daily. This essentially involves collecting measurement data from the Quantum Processing Unit (QPU), checking the state of the qubits, and readjusting them if necessary. If the measurements are unsatisfactory, the operating frequency must be tuned and further calibration would be necessary. Calibration is currently carried out using IQM software and by the company itself.
Dr. Xiaolong Deng: Under normal circumstances, calibration takes about 1,5 hours, though occasionally it may take several hours. This makes the operation of the quantum computer safer and more reliable. At our site, calibration is usually performed overnight. Since calibration can affect running jobs, or conversely be interrupted by jobs, we are currently developing rules and schedules for the process. Incidentally, understanding how jobs and calibration influence each other is an important research question for us. At present, however, there are only a few workloads on Euro-Q-Exa that run for several hours or longer than a day, and most jobs aren‘t affected by the nightly calibration.

How did that go?
Hossam: We still need more data and experience. However, we are already working with a large number of qubits, and quantum systems will continue to grow. As they do, increasing amounts of telemetry data are generated, along with data from cryostats and information about environmental and operational conditions that influence system performance. Through the LRZ monitoring tool Data Centre Data Base or for short DCDB, we have access to a broad data foundation that Ising and other AI models could analyse, classify, and enrich with contextual information. This would not only improve quantum computer performance but also help us evaluate conditions more accurately, make more informed operational decisions, and address additional research questions.
Xiaolong: For now, this remains an idea, but we could also build an agentic AI system based on Ising, meaning a system consisting of multiple AI models that analyse data, perform actions, evaluate the results, and continuously interact with the AI models. In addition to Ising, there are many other models for evaluating quantum computers and their operation. One could imagine developing an AI agent to answer user questions or optimise the system. Such an agent would access Ising, calibration data, and additional models, integrating their analyses into a seamless workflow. If users ask questions in natural language, the agent would process them and provide answers together with all relevant information. The implementation is technically feasible and appears practical for LRZ.

Error tolerance is another major technical challenge for quantum computers. Could AI help there as well?
Xiaolong: AI can help develop error-correction protocols. NVIDIA, for example, has an error-correction model based on Ising that uses neural networks. In simplified terms, hybrid quantum computing, where classical and quantum systems work together, involves an interplay between encoding and decoding. During encoding, classical data is transferred to qubits for computation. During decoding, the results produced by the quantum system are converted back into classical data. A supercomputer hosts a neural network for error correction. During decoding, data is sent there and checked. In this way, logical qubits can be derived from physical, or raw, qubits and used to measure computational results.
Hossam: In general, there are various approaches to increasing fault tolerance. In quantum error correction, a quantum circuit is designed to correct errors as they occur. We are also working on improving calibration, which is essentially a form of quantum error prevention. The goal is to prevent errors before they occur. This, in turn, enables fidelity thresholds to apply error correction, or fidelity. Error correction is a process involving many steps, and calibration is one of the most important steps for advancing quantum computing. Other factors include circuit quality, hardware, and programming.

What are the next tasks for the project and for the use of Ising and AI?
Xiaolong: In general we receive a great deal of feedback and many requirements from users. This enables us to answer questions in a targeted way and continue optimising the qubits. The next step is the automation of calibration. To achieve this, we first need complete control and understanding of the entire process. After that, we can integrate AI models such as Ising and combine them with our systems. This is another area in which we are collaborating with IQM and NVIDIA.
Hossam: By combining DCDB with Ising or other AI approaches, we can evaluate valuable data that is highly useful both for operational decision-making and for research. Interview: LRZ | vs