Skip to content

Architecture

Current repository structure

  • dfode_kit/cli/: CLI entrypoints and subcommands
  • dfode_kit/cases/: explicit case init, presets, sampling, and DeepFlame-facing helpers
  • dfode_kit/data/: contracts, HDF5 I/O, integration, augmentation, and labeling utilities
  • dfode_kit/models/: model architectures and registries
  • dfode_kit/training/: training configuration, training loops, registries, and preprocessing
  • docs/agents/: agent-facing operational and planning docs
  • tests/: lightweight repository and harness tests

Current refactor themes

1. Harness engineering

The repository now includes:

  • AGENTS.md
  • local verification commands
  • lightweight CI
  • documentation topology for agents and maintainers

2. Data contracts and workflow boundaries

A contracts layer is used to make HDF5 dataset assumptions explicit and testable. The canonical dfode_kit.data package now also owns the main data-preparation boundary:

  • HDF5 sampling outputs
  • HDF5-to-NumPy conversion
  • perturbation-based augmentation
  • CVODE/Cantera labeling
  • integration utilities used by downstream workflows

3. Config-driven training

The training stack is moving toward explicit config objects and registries so new model architectures and trainer types can be added without editing a monolithic training loop.

4. Agent-friendly CLI

The CLI now uses lighter command discovery and deferred heavy imports for improved usability in minimal environments.

Architectural end state of the recent refactor

The repository has now completed the transition away from the older compatibility layout. In particular, these legacy layers are removed from main:

  • dfode_kit/cli_tools/
  • dfode_kit/df_interface/
  • dfode_kit/data_operations/
  • dfode_kit/runtime_config.py
  • legacy dfode_core model/train compatibility packages

The current published docs should therefore treat cli, cases, data, models, runtime, and training as the only canonical implementation homes.