An Equivariant Machine Learning Decoder for 3D Toric Codes

This repository provides code for experiments using a Neural Decoder that works on 3D Toric codes.
The corresponding paper was published at: QCNC 2025.

Repository Structure

• models contains various decoder models and sub-packages with additional components.

  • auxiliar_components contains various components used in the NNs or decoding procedure.
  • pooling_layers contains adaptations of the GlobalAveragePooling layer.
  • residual_block contains different types of residual block implementations.

• src is a collection of classes and functions used in the experimentation.

  • data_analysis is a collection of functions for data analysis.
  • evaluation_metrics collects various evaluation methods for the decoding.

• config contains the configuration files used for experiments.

  • default contains default parameters for the experiments and API details to W&B.
  • net contains the specific parameters for the network architecture.