An AI-based decoder improves quantum error correction by reducing logical errors and accelerating decoding, while enabling researchers to build on the framework through publicly available code.
The developers have released the code and training assets for Ising Decoder ColorCode 1 Fast, an AI-assisted quantum error correcting decoder. The neural network performs pre-processing of error signals before feeding them to the Chromobius decoder, reducing logical errors and increasing decoding speed.
Simulation showed that the combination of both techniques reduced logical errors by a factor of 347.7 times and increased decoding speed by 7.3 times compared with using the Chromobius decoder alone. Testing was conducted on a quantum memory model with a code distance of 31 and a physical error rate of 0.3 per cent. The tests were performed with synthetic data and not with a real quantum computer.
The decoder includes a 17-layer three-dimensional convolutional neural network with 2.9 million parameters. The neural network functions as a pre-processing stage rather than an independent decoder. It analyses local error signals, filters unnecessary data, and sends a sparse error map to the Chromobius decoder.
The system focuses on colour-code quantum error correction, an approach that enables more efficient logical operations but introduces more complex error structures compared to traditional surface codes. This decoder aims to ease the computational load of the main decoding algorithm, the developers state that its advantages become more prominent as the code distance increases.
The decoder framework, training recipe, and other related materials have been made available under the Apache 2.0 license so that scientists could use, modify, and enhance the software. The decoder is part of the Ising family of AI models introduced earlier this year for quantum processor calibration and quantum error correction.
