Evaluating Machine Learning Approaches to Quantum Error Decoding

Overview

Modern Quantum computers are highly susceptible to errors and noise, and quantum error correction is necessary for scalable and reliable computation. Current classical decoders, such as Minimum-Weight Perfect Matching (MWPM), exist and can accurately correct errors, but can be slow in computation as code distances increase. This project compares its performance against a fully connected neural network (FCNN) to decode surface code errors, with the goal to achieve higher accuracy while reducing inference time. STIM was used to generate syndrome datasets for various code distances and physical error rates. An initial design was created and trained on labeled syndrome data. After training, it was evaluated in terms of logical error rate to physical error rate accuracy and inference time, with the training time also being documented. After testing, the model was then improved through alterations to its architecture, learning rate, and hardware efficiency (CPU vs GPU). It was evaluated again and compared against MWPM. First, the model was tested across different noise levels to determine its practicality in real scenarios, with it achieving over 90% accuracy on all tests. The initial model outperformed MWPM in accuracy for both distances 3 and 5. After iterations, the final model considerably increased efficiency in training epoch time (79% decrease) due to the change from CPU to GPU and modifications to its architecture. The final model’s inference time and accuracy was compared against MWPM, and it showed to maintain its higher accuracy while also having a significantly faster inference time. This data validates the solution: an FCNN decoder is a practical and more efficient alternative to classical decoding in tested conditions of small to medium scale code distances. The improvements in inference speed, training time, and accuracy suggest potential applications in real, current quantum error correction. In conclusion, this project demonstrates that machine learning models could serve as effective quantum error decoders, with having accuracy comparable to a classical method and improved computational efficiency, showing its capability to advance technology in quantum computing.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-37

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project