L2 Data Fabric Overview
The Data Fabric turns raw analog plasma telemetry into pristine, governed, AI-ready features for both machines, with lineage from feedthrough to feature store.
What L2 is for
Layer 2 sits between the microsecond control plane (L1) and the twin-and-AI layer (L3). Its single job is to convert the physical reality of a fusion shot into trustworthy numbers: it acquires 60+ analog channels, validates them, normalizes them into physics coordinates, engineers features, retains them at full fidelity, and serves them with provenance. Everything downstream — GNN imputation, PINN equilibria, anomaly ensembles, MPC — is only as good as this fabric.
The fabric is machine-agnostic by design. The same acquisition, validation, and governance layer runs for the breeder (Hyperion), a D-T spherical tokamak whose signals describe equilibrium, the negative-triangularity shape (delta -0.30) and disruption precursors, and for the burner, a D-3He tandem-mirror generator whose signals describe end-plug density, the ambipolar potential, and the direct-energy-conversion train.
The five stages
- Acquire: 60+ analog ports through cryo-rated vacuum feedthroughs at sub-microsecond sampling.
- Validate: high-frequency noise filtering, dead-sensor masking, calibration and drift correction.
- Normalize: raw channels mapped into flux/field coordinates common to the twin.
- Engineer: instantaneous Q, core pressure map, magnet strain deltas, quench precursors.
- Retain and serve: petabyte archive, low-latency feature store, RAG-ready vector database, strict lineage.
Design constraints
The fabric must be lossless where physics demands it (raw Mirnov and Thomson traces feed disruption forensics), lossy only where information theory permits, and always auditable. Every feature carries the identity of the channels, the calibration, and the transform that produced it, so any twin prediction can be traced to the exact bytes it rested on. These are design and simulation targets: construction of the breeder begins Q2 2027, with first-of-a-kind tritium near 2030.
This section documents the fabric end to end: the sensor front end, the acquisition chain, validation, normalization, feature engineering, the diagnostics constellation, storage, retrieval, and the governance and information theory that hold it together.