Machine learning for quantum error correction, control and calibration: Junction Brief – November 2026
Acentric Research is produced with AI tools under human review. The authors are not financial professionals, and nothing here is investment advice or a recommendation to buy or sell any security. AI analysis can be wrong; check every claim against its cited source.
This Brief is an initial scan, not a full investigation. The method and scores behind it are in AI × Quantum in the Landscape Report.
Overview
Quantum computers make mistakes. Their qubits, the quantum counterpart of bits, are disturbed by heat, vibration, stray electromagnetic fields and even cosmic rays. [1] The standard remedy is error correction: spread one piece of information over many physical qubits, check them constantly, and use a program called a decoder to work out which errors occurred. [1] This Brief covers artificial intelligence (AI) used for that decoding, and for tuning and controlling qubits.
What stands out is a clear laboratory result with a clear limit. In Google’s own tests, a neural-network decoder from Google DeepMind and Google Quantum AI made fewer errors than the decoders it was compared with on real data from a Google chip, but Google says it is too slow to correct errors on that kind of chip while the computer runs. [1][2] Federal funders have named the area and several projects have been chosen for award negotiations; funding is not a result. [3][4] The sources opened date from May 2024 to 2026, and the decoder result is from November 2024; later developments were not searched.
Current state
The decoder result. In a paper in Nature in November 2024, Google researchers describe a recurrent, transformer-based neural network that learns to decode the surface code, the leading error-correcting code. Its abstract states that it outperformed other state-of-the-art decoders on real data from Google’s Sycamore processor for distance-3 and distance-5 codes, and kept its advantage on simulated data up to distance 11. (Distance is the side length of the grid of qubits that protects one logical qubit; larger codes use more qubits.) [2][1] In the company’s blog post, the system, called AlphaQubit, made 6 percent fewer errors than a highly accurate but impractically slow method and 30 percent fewer than a fast, accurate one, in the largest Sycamore experiments. It was trained first on simulated data and then fine-tuned on thousands of real samples. The comparisons are Google’s own, on Google’s own chip. [1]
What was not searched. The decoder paper dates from November 2024. This Brief did not search for later decoder work, so developments since then are not covered. The comparison decoders in the post are algorithmic methods, not machine learning. [1]
The stated limit. The same post says that each consistency check on a fast superconducting chip is measured a million times a second, that AlphaQubit is “still too slow to correct errors in a superconducting processor in real time”, and that more data-efficient training is needed for machines with millions of qubits. [1]
Tuning and control. The 2025 review in Nature Communications, written mostly by NVIDIA staff, reports that AI agents built on language and vision models have been applied to calibrate quantum devices with performance comparable to human scientists, and surveys reinforcement learning, in which software learns by trial and reward, for qubit control. It lists limits: data collection and long training for reinforcement learning, and agents that cannot yet recover from all error cases. It adds that AI methods for reading out qubit states have not yet been shown for real-time, mid-circuit use, and that only simple models have run in real-time feedback. [5]
Readout. The review also reports that, on neutral-atom machines, convolutional neural networks (the image-recognition kind) reduced single-qubit measurement errors by up to 56 percent compared with traditional methods, and that neural networks have been applied to photon-count data from trapped ions to classify qubit states. [5]
A company in the area. Riverlane, a U.K. company, describes itself as building a quantum error correction stack with a decoder at its core. In May 2024 it announced a £2.1 million European Innovation Council grant to build a real-time decoder on a programmable chip (an FPGA). The announcement says that earlier decoding approaches cannot keep up with the volume of data from a quantum computer, which creates a backlog that slows and eventually stops the computation, and it does not say the decoder uses machine learning. [6] It is included here because it describes the speed requirement that any AI decoder would have to meet.
Public funding. DOE’s Genesis Mission opportunity includes a focus area on “AI for Control of Quantum System”, covering calibration, tuning, noise mitigation, error correction and readout, and another on using AI to find error correction for specific scientific applications. [3] The July 2026 list of projects chosen for negotiation includes titles on AI-driven discovery of error-correcting codes, AI-driven decoding for topological codes, and a project on decoding “at hardware speed”. [4]
How far it has come. A demonstrated laboratory result on real chip data, offline; real-time use is not shown in the sources opened.
Opportunities
Google’s researchers conclude that machine learning is “a strong contender for decoding in quantum computers”, and the company says its teams are working on the speed and data limits. [2][1] DOE’s call states that new AI approaches are needed for making real-time decisions under quantum uncertainty and for adapting to drift in real time. [3] The test is a decoder that is both accurate and fast enough to run alongside a live machine.
Challenges
- Physics and engineering. Google says each check on a fast superconducting chip is measured a million times a second, and Riverlane says a decoder that cannot keep up creates a backlog that stops the computation. [1][6]
- Scale. Real-data tests were on distance 3 and 5 codes; the larger sizes were simulated. [2]
- Training data. Google says more data-efficient training is needed for machines with millions of qubits. [1]
Data sources
- DOE Genesis Mission opportunity and the list of chosen projects, for federal programs in this area. [3][4]
- Research literature, through Europe PMC or arXiv, for decoder and control results. [7]
- Company announcements (Google, Riverlane), as company claims. [1][6]
- No registry covers quantum error correction; no coded record shows which decoders use machine learning.
Changelog
- 2026-11-18: First published.
Sources
- Google DeepMind and Google Quantum AI, “AlphaQubit tackles one of quantum computing’s biggest challenges”, Google blog, 2024-11-20 (company statement). blog.google (accessed 2026-11-18)
- Johannes Bausch and others, “Learning high-accuracy error decoding for quantum processors”, Nature 635, 834–840, published 2024-11-20. doi.org (abstract read through the Europe PMC API, record 39567694, accessed 2026-11-18; the full text was not read)
- US Department of Energy, “The Genesis Mission: Transforming Science and Energy with AI”, Notice of Request for Application DE-FOA-0003612, issued 2026-03-17, Amendment 000003 of 2026-04-20; challenges 7 and 8. science.osti.gov (accessed 2026-11-18)
- US Department of Energy, “Energy Department announces first Genesis Mission projects to accelerate new era of AI-driven scientific discovery”, list of selections for award negotiations, posted 2026-07-22 (the list states that selection is not a commitment to issue an award or provide funding). energy.gov (accessed 2026-11-18)
- Yuri Alexeev and others, “Artificial intelligence for quantum computing”, Nature Communications 16, 10829, published 2025-12-02 (a review; the first author and most listed authors are at NVIDIA, which sells AI computing hardware). doi.org (full text read through the Europe PMC API, PMC12673084, accessed 2026-11-18)
- Riverlane, “Riverlane awarded £2.1m by Horizon Europe to develop the next generation of its quantum error correction decoder”, press release, 2024-05-01 (company statement). riverlane.com (accessed 2026-11-18)
- Europe PMC, REST API (search and full-text endpoints), queried 2026-11-18. ebi.ac.uk