# Machine-learning decoders for fusion-based quantum computing in real time

Quantum.world

> Transformer decoder matches integer-programming accuracy while running fast enough for real-time photonic fusion QEC at erasure rates near threshold.

**arXiv**: 2611.13630

**Published**: 2026-11-10

**Authors**: Daniel Barter, Jacob C. Bridgeman, Chris Dawson, Terry Farrelly, Shane Fleming, Mercedes Gimeno-Segovia, Paul Hartke, Timothy Hein, Sean Howell, Gregory Hyegang Jun, Siddhartha Kamalakara, Yehua Liu, Jose M. Monsalve Diaz, Eddie Richter, Sam Roberts

## AI summary

## Overview

**Research area:** Quantum computing — specifically fault-tolerant quantum error correction, machine-learning decoders, and real-time classical control hardware for photonic fusion-based quantum architectures.

**Technical level:** Intermediate. The abstract assumes familiarity with error correction terminology (surface codes, erasure errors, threshold, logical error rate), but its central claims — a transformer-based decoder that is both accurate and fast enough for real-time use on an FPGA — are accessible to anyone with a working knowledge of quantum error correction.

**Scope:** The paper introduces and benchmarks PRISM, a transformer-based machine-learning decoder for erasure-dominated fault-tolerant quantum computing, and reports an optimized real-time FPGA implementation with microsecond-scale latency.

## What This Paper Is About

Fault-tolerant quantum computers need a classical "decoder" — software or hardware that looks at error-prone measurements of physical qubits and infers what went wrong — and that decoder must be accurate, fast, and scalable all at once. In photonic architectures built on fusion, photon loss and probabilistic fusion measurements produce *erasures* (errors whose location is known), which is a different error profile from what most decoders were designed for. The paper asks whether a machine-learning decoder can exploit that known-location information to match the best conventional decoders in accuracy while running fast enough to keep up with a real quantum machine.

## Key Contributions

1. **PRISM, a transformer-based ML decoder:** A decoder designed specifically to exploit the known locations of erasures in protocols that suffer from both Pauli errors and erasures — a mix characteristic of photonic fusion-based hardware.
2. **Broad benchmarking across code families:** Evaluation on fault-tolerant protocols based on surface codes, color codes, and quantum low-density parity-check (LDPC) codes, at code distances up to 24 — indicating the decoder is not tied to a single code family or a single small-scale setting.
3. **Accuracy competitive with integer programming:** PRISM is reported to match the logical error rates of integer-programming-based decoders (typically treated as a high-accuracy reference) while being significantly faster, including in the demanding regime of error rates up to 50% of threshold.
4. **A real-time hardware implementation:** Co-optimization of the decoder and its hardware realization on an AMD VP1902 FPGA, targeting the latency constraints of live quantum error correction.

## Main Findings

- **Accuracy matches a strong classical baseline:** Across the surface-code, color-code, and quantum LDPC benchmarks, PRISM's logical error rates match those of integer-programming-based decoders. The abstract does not report the specific error-rate values or the number of samples behind them.
- **Substantially faster than integer programming:** PRISM is described as significantly faster than the integer-programming decoder it matches in accuracy. The abstract states the qualitative result but does not give a speedup factor.
- **Holds up near threshold:** The accuracy and speed advantages are reported to persist at physical error rates up to 50% of threshold — a regime where decoders typically struggle, and one made harder by the high erasure rates of photonic hardware.
- **Microsecond-scale latency on FPGA:** The optimized implementation on an AMD VP1902 FPGA decodes logical blocks of distance 8 in both space and time with a latency of 6.59 microseconds.
- **Amortized throughput per block:** An amortized decoding time of 1.56 microseconds per logical block is reported as achievable with that implementation.
- **Overall claim:** The results are presented as evidence that machine-learning decoding can deliver accurate, low-latency inference for erasure-dominated fault-tolerant quantum architectures — that is, ML decoding is viable as real-time control hardware, not just an offline analysis tool.

## Methodology in Plain English

The team built a decoder around a transformer — the same broad class of neural network architecture behind modern language models — and trained it to take the syndrome data from error-correcting codes and infer the likely underlying error. The key design choice is that the decoder is told, as part of its input, *where* erasures occurred, rather than being forced to deduce those locations from scratch; in photonic hardware, photon loss and failed fusion measurements reveal this directly. To test generality, they ran the same decoder against many different fault-tolerant protocols spanning three code families (surface, color, and quantum LDPC) and a range of code sizes. They compared its logical error rate against an integer-programming decoder, which solves the decoding problem via optimization and serves as an accuracy reference. Finally, they tuned both the decoder and its hardware mapping together to fit within the tight timing budget of a real-time control loop, and measured latency and per-block decoding time on an FPGA rather than on a general-purpose CPU or GPU.

## Why This Matters

Decoding is one of the practical bottlenecks standing between fault-tolerant quantum computing on paper and a working machine. Quantum architectures produce syndromes far faster than a naive software decoder can process them, and a decoder that is accurate but too slow, or fast but inaccurate, is unusable. This work targets both constraints at once and, importantly, does so for the erasure-heavy error model of photonic fusion-based hardware — a platform where erasures dominate and where decoders designed for depolarizing noise may be poorly matched. Showing that a learned decoder can match an optimization-based reference across three code families, and then run in microseconds on an FPGA, is a concrete step toward treating ML decoding as production control hardware.

Real-world applications:

- **Photonic quantum computing systems:** Real-time syndrome decoding for fusion-based photonic machines, where photon loss makes erasure handling central.
- **Quantum error-correction control stacks:** Low-latency decoders as a component in the classical control layer of any fault-tolerant quantum processor.
- **Hardware accelerator design:** FPGA (and eventually ASIC) implementations of neural-network decoders, informing how quantum control electronics are architected.
- **ML for scientific instrumentation:** The latency-optimization and hardware-mapping techniques transfer to other settings requiring learned inference inside tight real-time feedback loops.

**Industry relevance:** The work sits at the intersection of quantum hardware companies, FPGA/accelerator vendors (the AMD VP1902 is named explicitly), and the emerging quantum-control-electronics supply chain. Companies building photonic quantum computers have a direct stake in erasure-aware, microsecond-latency decoding, and the result suggests that ML-based decoders can be evaluated on the same latency metrics as conventional ones rather than only on accuracy.

## Future Directions

- **Extending to larger distances and longer times:** The hardware result is reported for distance 8 in both space and time; scaling the real-time implementation to larger logical blocks while holding latency is the obvious next test, and the abstract does not say how latency scales.
- **Closing the gap to optimal decoding:** Matching integer programming is a strong result, but whether a learned decoder can exceed it — particularly near threshold and at high erasure rates — remains open.
- **Generalization beyond the training distribution:** The abstract does not detail how PRISM behaves on codes, noise models, or erasure rates it was not trained on, which matters for deploying a decoder on real, drifting hardware.
- **Comparison against other fast decoders:** The benchmarks compare against integer programming; comparisons with other low-latency decoders (e.g., dedicated real-time decoders not based on ML) are not described in the abstract, leaving the competitive landscape partly unstated.

## Target Audience

Quantum error-correction researchers and engineers, particularly those working on photonic and fusion-based architectures; hardware architects designing real-time control electronics for quantum processors; and machine-learning researchers interested in deploying learned models under hard latency constraints. The paper is also useful to technical decision-makers assessing whether ML decoding is ready for integration into a quantum computing stack, though readers without error-correction background will want to review surface codes, erasures, and thresholds first.

## Authors’ abstract

Fault-tolerant quantum computation requires decoders that combine accuracy, throughput, low latency, and scalability to large numbers of physical qubits. In photonic fusion-based architectures, erasures from photon loss and probabilistic fusion measurements are a dominant error mechanism, motivating decoders that remain accurate at high erasure rates. We introduce PRISM, a transformer-based machine-learning decoder designed to exploit the known locations of erasures in…

[The full abstract on arXiv](https://arxiv.org/abs/2611.13630)

[Read the original paper](https://arxiv.org/abs/2611.13630)

---

Canonical: https://quantum.world/paper/machine-learning-decoders-for-fusion-based-quantum-computing-in-real-2611.13630
