QEC lake: a data lake for quantum error correction

Quantum error correction (QEC) is what makes a fault-tolerant quantum computer possible: physical qubits are noisy, so QEC codes spread logical information across many physical qubits and use repeated syndrome measurements to detect and correct errors. Decoding — turning a stream of noisy syndrome measurements into “which qubits flipped” — is increasingly done with AI models rather than hand-designed algorithms, because learned decoders can adapt to the messy, correlated noise of real hardware better than classical decoders can.

The catch: training a good decoder needs a lot of QEC data, and that data is a mess. Every code family (surface codes, color codes, …) and every hardware platform emits data in its own schema. Datasets are scattered across incompatible repositories, often undocumented, and rarely large enough on their own. Comparing two decoders fairly is hard when they were never trained or evaluated on comparable data in the first place.

QEC lake, the paper Max de Groot, Andrei Ilinescu, Aditya Shankar, Yuandou Wang, Wenbo Sun, Floris Geerts, Rihan Hai and I demoed at VLDB 2026, is a domain-specific data lake built to fix exactly this. It:

• Continuously crawls and refreshes open QEC datasets from across code families and hardware platforms into one place.
• Generates new data — user-configurable simulated circuits, and synthetic variants derived from real hardware data — to fill in the gaps where real data is scarce.
• Manages metadata for discovery, so a researcher can actually find the subset of data relevant to their code family, distance, and noise model instead of digging through disconnected repos.
• Prepares model-ready inputs, converting heterogeneous raw data into the tabular formats tree-based models and tabular foundation models expect, plus an AI-assisted natural-language interface for exploring the data.

At the VLDB demo, attendees ingested sources, generated new datasets on the spot, explored them through structured and natural-language queries, and exported model-ready training data — the full pipeline from “scattered QEC data” to “something a decoder can train on,” in one sitting.

To make the underlying decoding problem tangible for a general audience (not everyone at a database conference has a quantum information background), I also built QEC Tiles: Stabilizer Rush, a small browser game where you play the decoder in real time.

Paper, abstract, and PDF are on the publications page (Patwardhan et al., 2026).