Interpolator with Mass Conservation Logic

[roxas6path]

Add mass-conservation to Time Interpolator Runner and align I/O with stand-alone inference

1) Mass conservation in Time Interpolator Runner

Adds the necessary optional mass-conservation logic to the Time Interpolator Runner.

• If the forecast being interpolated includes accumulated-from-start fields, this work may be affected by feat: abstracting accumulation anemoi-datasets#326

---

2) Approach to suit stand-alone inference - run independently from forecaster

Aiming to implement this in a way that works cleanly with prepML which is slightly different to the initial design that was implemented

• a) Other models typically request data once for a single set of initial conditions.

• b) The Time Interpolator requires multiple sets of initial conditions—the temporal boundaries for each interpolation window (e.g., t=0 and t=24 for the first 24-hour window, then t=24 and t=48 for the second).

• c) Proposed method:

  • Have PrepML request all required input data up front and write it to local storage.
  • Have anemoi-inference make data-access calls only to this local dataset, rather than issuing repeated requests to MARS or a single batched MARS call.

---

Acceptance criteria

• Interpolated outputs satisfy mass-conservation checks across each window.
• Accumulated-from-start fields are handled correctly when present.
• PrepML can prefetch and persists all required inputs for each interpolation window.
• Backward-compatible configuration flags (opt-in) and clear documentation.
• Add necessary prepml config updates so that variable naming is handled e.g. in training the input variables for accum are named tp_accum, cp_accum for example, whereas in inference they would be just 'tp' and 'cp'

---

Related PR(s)/Branches:
https://github.com/ecmwf/anemoi-inference/tree/feature/326-intrp-mass-conservation

Describe the solution you'd like

No response

Describe alternatives you've considered

No response

Additional context

No response

Organisation

No response

[roxas6path]

notes on whats implemented

• Ability to run on forecasts w/ accumulations (class class MarsFc
• Ability to interpolate reanalysis, w/ accumulations sourced from related forecasts (class MarsFcAlignToInits)
• Compatibility w/ ensemble interpolator
• Compatibility w/ interpolators that output the final n'th value instead of leaving it to be imputed e.g. explicit_target_indexes=[1,2,3,4,5,6] and explicit_input_indexes = [0,6]