Heat Transfer Recipe¶

This recipe trains neural surrogate models for heat transfer simulations on the SIMSHIFT Heatsink benchmark dataset.

We published a report on baseline models on W&B

SIMSHIFT Heatsink Example

Figure: Visualization showing the unnormalized temperature error of a trained ABUPT model.

Overview¶

Given a heatsink geometry described by simulation parameters (number of fins, fin dimensions, inlet temperature, etc.), the model predicts the full 3D volume fields:

Velocity (3 components)
Temperature (scalar)
Pressure (scalar)

The recipe includes:

Dataset: SimshiftHeatsinkDataset – 819 heat transfer simulations with varying fin configurations
Models: AB-UPT and Transformer architectures, both conditioned on simulation parameters
Pipeline: Volume point sampling with configurable anchor subsampling
Callbacks: Denormalized RMSE/nRMSE metrics and interactive 3D plotly visualizations
Normalization: Position scaling to [0, 1000], mean-std normalization for all other fields

Running an experiment¶

All commands must be run from the recipes/heat_transfer/ directory.

Local training¶

uv run noether-train \
  --hp configs/train_simshift_heatsink.yaml \
  +experiment/simshift_heatsink=ab_upt \
  tracker=disabled \
  dataset_root=/path/to/simshift_heatsink

SLURM submission¶

uv run noether-train-submit-job \
  --hp configs/train_simshift_heatsink.yaml \
  +experiment/simshift_heatsink=ab_upt \
  dataset_root=/path/to/simshift_heatsink

Available experiments¶

Experiment config	Model	Description
`+experiment/simshift_heatsink=ab_upt`	AB-UPT	Anchored-Branched UPT with 5 physics blocks
`+experiment/simshift_heatsink=transformer`	Transformer	Standard Transformer

Project structure¶

recipes/heat_transfer/
├── callbacks/          # Evaluation and visualization callbacks
│   ├── heat_transfer_metrics.py          # RMSE and nRMSE in physical space
│   └── heat_transfer_visualization.py    # 3D plotly scatter plots (GT, prediction, error)
├── configs/
│   ├── experiment/simshift_heatsink/     # Per-model experiment overrides
│   ├── model/                            # Model architecture configs
│   ├── slurm/                            # SLURM job configuration
│   ├── tracker/                          # WandB / disabled tracker configs
│   └── train_simshift_heatsink.yaml      # Main training config
├── model/
│   └── heat_transfer_transformer.py       # Transformer wrapper for volume-only + conditioning
├── pipeline/
│   └── heat_transfer_pipeline.py         # Volume point sampling and anchor creation
└── README.md

Configuration¶

Key parameters in train_simshift_heatsink.yaml:

Parameter	Default	Description
`dataset_root`	–	Path to the SIMSHIFT Heatsink dataset (required)
`pipeline.num_volume_points`	4096	Number of volume points sampled per training sample
`trainer.max_epochs`	75	Total training epochs
`trainer.precision`	float16	Training precision
`optimizer.lr`	1e-3	Peak learning rate (cosine decay with 5% warmup)

Callbacks¶

Training automatically logs:

Validation loss every epoch (OfflineLossCallback)
Denormalized RMSE and nRMSE per field on val and test sets (HeatTransferMetricsCallback)
3D scatter visualizations every 10 epochs on the first test sample (HeatTransferVisualizationCallback) – ground truth, prediction, and per-point relative L2 error