Low-field MRI digital twins¶

Adelpha MRI’s twin is DTAM, vendored at sidecar/dtam/. The GUI is a read-only observer of Twin HTTP state plus explicit forecast actions. It is built for low-field magnets (bundled 48 / 47 / 64 mT Halbach, Delta v2, simulated scanner) rather than clinical 1.5 T / 3 T systems.

Adelpha architecture: Tauri shell, Python runtime, DTAM twin, and MRI services. — Desktop supervisor mounts DTAM next to the MRI façade. Twin telemetry never crosses into TypeScript hardware I/O.

Provenance is the product¶

Every live number in the Digital Twin workspace is labeled measured, estimated, predicted, or nominal. A PINN forecast must not look like a thermocouple.

Source	Meaning in the UI
`measured`	Sensor or adapter reading
`estimated`	Twin estimate from the current window
`predicted`	Horizon forecast only (`POST /twin/forecast`)
`nominal`	Design value (for example packed \(B_0\))

The client polls GET /twin/state without a horizon so live channels stay cheap. Predicted scalars stay null until the operator requests a horizon. Mapping: systemStateToTelemetry() in the GUI; nested SystemState in the side panel (state.thermal?.mean_magnet_temperature_c?.value, …).

Physics slice that actually runs¶

DTAM Phase 2 is a working thermal → \(B_0\) vertical slice, not a full Maxwell solver:

Simulated (or measured) temperatures
  → synchronization window
  → thermal state estimate
  → thermal→B0 model
  → B0 / f0 estimate (+ optional prediction)

Coupling used in the twin:

\[ \Delta B_0(t) = \alpha_T \, \Delta T(t) \]

\[ f_0(t) = \frac{\gamma}{2\pi}\, B_0(t) \]

Default \(\alpha_T\) is \(-5\times10^{-5}\) T/°C in configs/models.yaml. Resonant frequency is reported in MHz; temperature in °C.

Optional thermal PINN: if model.pt / model.onnx exists, ThermalMagneticTwin.update(..., predict_horizon_s=...) uses it; otherwise prediction falls back to a constant heating rate. Train with uv sync --extra pinn inside sidecar/dtam/ — that is a developer path, not bundled in the installer.

EMI and RF channels are heuristic estimators, not a certified EMC report. Treat them as twin state, not lab accreditation.

Scanner adapters¶

Settings → Digital Twin chooses the Python adapter the supervisor starts:

Profile	Role
Simulated scanner	Plant + drift scenarios without hardware
48 mT Halbach	Bundled low-field identity for the twin

The 3D Model page is a separate concern: which GLB you see. You can view Delta v2 CAD while the twin still runs the simulated adapter. Restart the Python runtime after a twin profile change.

Authoritative twin math and package map: sidecar/dtam/docs/digital_twin/ in this repository (not published on the Adelpha site). Operator surface: Dashboard and Twin API.