LASERNet

Spatiotemporal deep learning for predicting microstructure evolution in laser-based additive manufacturing.

Highlights

• Multiple architectures (CNN-LSTM, PredRNN, MLP baselines)
• End-to-end training/evaluation/prediction scripts
• Reproducible experiments via YAML configs

Installation

This project uses uv and Python 3.12.

uv sync --locked --dev

Download dataset locally

The dataset is hosted at https://storage.googleapis.com/lasernet-data. Download it into a relative ./data folder from the project root.

Use wget to download the files over HTTPS:

mkdir -p data && cd data
wget -i https://storage.googleapis.com/lasernet-data/filelist.txt

Data setup on HPC

Data and models uses large amount of storage space and is therefore stored on a seperate blackhole scratch drive (≈50 GB) on DTU HPC. Thus, it is important to make symlinks to data/ and models/ folders, so that they are relative to the project root directory.

ln -s "$BLACKHOLE/models" /zhome/b0/7/168550/Github/LASERNet/models
ln -s "$BLACKHOLE/data" /zhome/b0/7/168550/Github/LASERNet/data

Quickstart

Train, evaluate, and predict via CLI scripts:

uv run src/lasernet/train.py --network deep_cnn_lstm_large --field-type microstructure
uv run src/lasernet/evaluate.py --network deep_cnn_lstm_large --field-type microstructure
uv run src/lasernet/predict.py --network deep_cnn_lstm_large --field-type microstructure --timestep 21

Experiments from YAML configs:

uv run src/lasernet/experiments/experiments.py

Results

Model prediction comparison (timestep 21). Qualitative comparison across selected models for a fixed timestep.

Timestep evolution (18 → 21). Predicted temporal evolution in the microstructure field.

MSE error maps. Spatial error distribution highlighting challenging regions.

Tests

uv run pytest -q

Project structure

├── .github/                  # CI workflows
├── configs/                  # Experiment configs
├── data/                     # Data (raw/processed)
├── models/                   # Model checkpoints
├── notebooks/                # Demos and exploration
├── results/                  # Evaluation artifacts
├── src/lasernet/             # Package source
│   ├── data/                 # Dataset + normalization
│   ├── models/               # Model implementations
│   ├── evaluate.py           # Evaluation CLI
│   ├── predict.py            # Prediction CLI
│   ├── train.py              # Training CLI
│   └── utils.py              # Utilities
└── tests/                    # Unit tests