A Python toolkit to estimate the front-view width of a liquid drop sliding down an inclined surface using only side-view measurements.
Implementation of Shumaly et al., Scientific Reports (2024).

Why is this useful?
Drop width enters directly into both the lateral adhesion (Furmidge) and dynamic friction force equations.
Conventional approaches require a second camera or mirror that shrinks the observable area and complicates alignment.
Our model keeps the set-up minimal while still giving width data for the entire 5 cm track—even when the drop interacts with surface defects.

• Features
• Installation
• Usage
  • Training
  • Inference
• Data Structure
• Model Architecture
• Output
• Project Structure
• Requirements
• License
• Citation
• Contact

• Train an LSTM-based regression model on side-view drop features.
• Predict drop width from new video data for any Video ID.
• Automatic normalization, sliding-window slicing, and result plotting.
• Example scripts: train_model.py and inference.py.

git clone https://github.com/AK-Berger/Drop_width_estimation.git
cd Drop_width_estimation
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

python train_model.py \
  --data-file data/dataset.xlsx \
  --model-output models/lstm_weights.h5 \
  --epochs 2500 \
  --batch-size 16 \
  --window-size 20

• --data-file: Path to the Excel dataset (must include status).
• --model-output: Path to save trained model weights (.h5).
• --epochs, --batch-size, --window-size: Training parameters.

python inference.py \
  --data-file data/dataset.xlsx \
  --weights-file models/lstm_weights.h5 \
  --video-id 14 \
  --window-size 20 \
  --output output/14.png

• --data-file: Source Excel (with Video ID, features, status).
• --weights-file: Trained LSTM weights (.h5).
• --video-id: Integer ID of the video to process.
• --window-size: Sliding window length (frames).
• --output: (Optional) Path to save plot (PNG). Without this, the plot displays interactively.
• Results Excel: Saved automatically to output/<datafile_stem>_<video_id>.xlsx.

• dataset_<video_id>.xlsx: Filtered data for each Video ID (e.g., dataset_14.xlsx, dataset_234.xlsx).

• LSTM layer: 48 units, tanh activation, L2 regularization (λ=0.01)
• Dropout: 0.5
• Dense output: 1 unit, tanh activation
• Input shape: (window_size, 6) features per frame
• Loss: MSE, Optimizer: Adam
• Label scaling: Ground-truth width scaled from cm to μm

• Excel: output/<datafile_stem>_<video_id>.xlsx containing original features plus:
  • Estimated Width (μm)
  • Drop width (cm) (measured)
• Plot: Overlaid estimated vs measured widths vs frame index

Drop_width_estimation/
├── data/
│   └── dataset.xlsx
├── models/
│   └── lstm_weights.h5
├── output/
│   ├── dataset_14.xlsx
│   └── dataset_234.xlsx
├── src/
│   └── drop_width/
│       ├── preprocessing.py
│       ├── model.py
│       └── side_to_width.py
├── train_model.py
├── inference.py
├── requirements.txt
├── The setup.png
└── README.md

See requirements.txt for full package versions.

Licensed under the GNU General Public License. See LICENSE for details.

@article{Shumaly2024,
  title   = {Estimating sliding drop width via side-view features using recurrent neural networks},
  author  = {Shumaly, Sajjad and Darvish, Fahimeh and Li, Xiaomei and Kukharenko, Oleksandra and Steffen, Werner and Guo, Yanhui and Butt, Hans-J{"u}rgen and Berger, R{"u}diger},
  journal = {Scientific Reports},
  volume  = {14},
  pages   = {12033},
  year    = {2024},
  doi     = {10.1038/s41598-024-62194-w}
}

• Technical: Sajjad Shumaly — shumalys@mpip-mainz.mpg.de