Combustion Flame Temporal Super-Resolution

This repository provides a complete training, inference, and evaluation framework for combustion flame-field temporal super-resolution on RealPDEBench combustion/hf_dataset/real.

The task setup is fixed to 10-frame windows:

• input conditions: frame 0 and frame 9
• targets to reconstruct: frames 1..8
• method: frame autoencoder + conditional latent DiT flow matching

The framework is prepared for later training, but no training jobs are launched by default.

Quick Start

1. Check that the dataset root exists:

python scripts/inspect_combustion_data.py --config configs/autoencoder_base.yaml

2. Train the autoencoder later when ready:

torchrun --standalone --nproc_per_node=1 scripts/train_autoencoder.py --config configs/autoencoder_base.yaml

3. Train the flow matching model after filling in the autoencoder checkpoint:

torchrun --standalone --nproc_per_node=1 scripts/train_flow_matching.py --config configs/flow_matching_base.yaml

4. Run interpolation inference:

python scripts/infer_interpolation.py \
  --config configs/flow_matching_base.yaml \
  --flow-checkpoint results/flow/combustion-flow-v0/checkpoints/best.pt \
  --sample-index 0

5. Run evaluation:

python scripts/evaluate.py \
  --config configs/flow_matching_base.yaml \
  --flow-checkpoint results/flow/combustion-flow-v0/checkpoints/best.pt

Cluster Entrypoints

Two cluster-style launchers are included and follow the same environment-variable pattern as the referenced GLD script:

• entrypoints/entry_train_autoencoder.sh
• entrypoints/entry_train_flow_matching.sh

Supported environment variables include EXP_NAME, CONFIG_PATH, RESULTS_DIR, LOG_ROOT, PRECISION, RESUME, CKPT_PATH, WANDB_ENABLED, PROJECT, ENTITY, and distributed launch variables such as NPROC_PER_NODE, NNODES, NODE_RANK, MASTER_ADDR, and MASTER_PORT.

Both entrypoints launch training with torchrun, and both also accept arbitrary CLI overrides as trailing key=value arguments.

Documentation

See docs/usage.md for the full workflow, configuration guide, DDP launch examples, and output structure.