Analysis And CNN

analysis/BGK1D

The analysis package provides benchmark, convergence, and cross-scheme utilities together with plotting helpers.

CNN/BGK1D1V

The CNN code is currently organized into five layers.

  • util/: model architecture, input-state encoding, and loss definitions
  • gen_traindata_v1/: legacy NPZ-based data pipeline
  • gen_traindata_v2/: shard-based PT pipeline with dataset and case manifests
  • train/: training and sweep entry points
  • evaluation/: cached warm-evaluation engine and compatibility wrappers

The current train.py also exposes conv_type and forwards it into the train-time warm-evaluation implicit configuration.

Model and data conventions

  • MomentCNN1D is a lightweight residual 1D CNN with an optional gated tail head.
  • Inputs may be primitive (nut) or conservative (nnuT) and may optionally include temporal history via prev_delta.
  • Targets are either dw = (\Delta n, \Delta u, \Delta T) or dnu = (\Delta n, \Delta \nu, \Delta T).
  • The implicit BGK1D stepper loads checkpoints and uses the predicted moment increments to initialize Picard iteration.