AQ1-Emerald-d3

Reimplementation of AlphaQubit (Bausch et al., Nature 2024) trained on d=3 rotated surface code data and finetuned on real IQM Emerald QPU hardware.

Architecture

Recurrent transformer decoder: StabilizerEmbedding → SyndromeTransformer (4 layers) → GRU → Readout

• Parameters: 1,117,185

• Code: d=3 rotated surface code, Z-basis memory

• Hardware: IQM Emerald (54-qubit square lattice)

Training

• T1: 2.16M Stim simulator samples, 120 configs, 20 epochs, lr=1e-4

• T3: 12k real IQM Emerald QPU shots (r=1), lr=2e-6, EMA=0.999, 20 epochs

Performance

Evaluated on held-out real IQM Emerald QPU data (4k shots, r=1):

• QPU LER (r=1): 0.446

• Random baseline: 0.500

• MWPM baseline: 0.393

Note: IQM Emerald d=3 is above the error correction threshold.

Usage

import torch

from huggingface_hub import hf_hub_download

from model import AQ1Decoder



ckpt_path = hf_hub_download("RafalMa/AQ1-Emerald-d3", "tier3_best.pt")

model = AQ1Decoder(n_stabilizers=8, d_model=128, n_heads=4, n_transformer_layers=4)

ckpt = torch.load(ckpt_path, map_location='cpu')

model.load_state_dict(ckpt['model_state'])

model.eval()

Citation

@article{bausch2024alphaqubit,

  title={Learning high-accuracy error decoding for quantum processors},

  author={Bausch, Johannes and others},

  journal={Nature},

  year={2024}

}