Machine learning quickly and accurately diagnoses noise sources in quantum computers.
Researchers at the Indian Institute of Technology Madras have developed a machine learning technique that can quickly identify noise sources in quantum computers. By training an artificial neural network on simulated data and testing it on IBM’s quantum processors, the researchers showed that they could more accurately diagnose disturbances and design targeted suppression methods.
“We leverage artificial neural networks trained on well-designed synthetic data to rapidly predict noise features with minimal loss of accuracy,” study co-author Professor Siddharth Donkar said in an email.
The promises and problems of quantum computing
Quantum computers are often touted as the next big leap in technology. Unlike regular computers, which use bits (small switches that are either 0 or 1), quantum computers use qubits and can be in multiple states at the same time. This ability provides significant advantages in certain types of calculations, from designing new materials to cracking codes that even the fastest supercomputers have difficulty with.
The problem is that qubits are fragile. They rely on delicate quantum effects that can disappear with the slightest stimulation from the outside world. Dhomkar said, “Anything that can interact with a qubit can potentially destroy the quantum coherence (a degree of quantumness) that is essential to the operation of a quantum computer. The largely uncontrollable interaction of a qubit with its surrounding environment creates so-called dephasing noise.”
Researchers have long sought ways to protect qubits from interference. But the first step is figuring out exactly where the noise is coming from, which turns out to be difficult. Disturbances can change over time, and measuring them directly is time-consuming and complex.
“Deciphering the precise nature of these complex interactions requires the implementation of time-consuming and complex quantum protocols,” Domker says. When scientists try to measure disturbances, they often get average images that omit important details. As a result, many strategies for shielding qubits remain difficult to implement.
Machine learning solutions
To break out of this impasse, Dhomkar and his colleagues turned to artificial intelligence. Their approach is inspired by the same techniques used in other fields, where computers learn to identify cats and faces by being shown thousands of examples. The researchers created extensive simulation data that shows how qubits are perturbed by their environment. Once the computer “learned” these patterns, it could quickly find the same signature in real experimental data.
“Neural network architectures are derived from models commonly used for image recognition and processing tasks,” Dhomkar explains. The result is speed. Instead of spending weeks running complex tests to figure out what’s messing up a qubit, a machine learning system can come up with an answer in a fraction of the time.
The research team tested their method on IBM’s superconducting quantum processor. These devices use tiny electrical circuits cooled to near absolute zero and operate like qubits. In this state, electricity flows without resistance, allowing circuits to create and maintain quantum states that are weak enough to be useful for calculations.
“We use this methodology to characterize different IBM qubits, show the temporal evolution of the underlying noise, and in principle build customized sequences that can help suppress it,” Dhomkar said.
The results were promising. By diagnosing disturbances more quickly and accurately, researchers may be able to suggest ways to improve the performance of qubits. “We have already implemented the protocol on IBM qubits, and the plan is to use this technology to benchmark and compare superconducting qubits being studied in different laboratories around the world,” Dhomkar said. “This could provide valuable insights to improve manufacturing strategies, thereby increasing the quality of qubits.”
Beyond one type of qubit
Although this study focuses on superconducting qubits, the researchers believe their method can be used for other designs as well. “While the technique originally developed here is hardware agnostic, the current implementation was targeted at transmon qubits,” Dhomkar says. The key is to model the environmental perturbations for each type of qubit. “We have already implemented a similar strategy in optical spin systems and can indeed extend it further.”
This flexibility means this approach could help advance an entire field that is still experimenting with many competing technologies.
Researchers don’t stop here. “We are currently developing ways to deal with more complex noises,” says Dhomkar. Simply put, they are working on ways to deal with more complex and unpredictable types of disturbances. They are also exploring new forms of artificial intelligence to proactively design better ways to control quantum computers.
“We are also exploring new AI techniques that allow us to design customized ways to perform quantum operations more efficiently, even when the hardware is not perfect.”
Quantum computers still have a long way to go, but this research represents a promising step forward. By teaching machines to recognize and counter the hidden obstacles that plague qubits, researchers are finding new ways to bring the dream of quantum computing closer to reality.
References: Bhavesh Gupta et al. Fast noise spectroscopy of transmon qubitsAdvanced Quantum Technologies (2025). DOI: 10.1002/qute.202500109
Feature image credit: Gerd Altmann (via Pixabay)
