TransMICRO-Net

TransMICRO-Net — Transformer for Multi-material Inter-system Cross-mechanism Representation and Outcome

A quasi-multimodal, multi-task deep learning framework that simultaneously reconstructs full stress-strain curves and predicts peak stress, yield point, and toughness of biomaterial hydrogels from limited compression data.

This repository accompanies the manuscript "TransMICRO-Net: A Quasi-multimodal Multi-task Interpretable Attention-based Deep Learning Framework for Accurate Prediction and Cross-mechanism Deciphering of Mechanical Fingerprints in Biomaterial Hydrogels" (Frontiers). The full case-study analysis and additional figures are provided in the Supplementary Material of that article.

---

Highlights

• Quasi-multimodal inputs — 200-point stress sequence + material identity + concentration, conditioned via FiLM layers.
• Convolutional residual backbone + multi-head self-attention for long-range strain-axis dependencies.
• Uncertainty-weighted multi-task loss (Kendall et al. 2018) balancing curve reconstruction and property regression.
• Snapshot ensembling + Mixup + Gaussian noise augmentation for the data-scarce regime (18 training specimens, 3-fold CV).
• Interpretability toolbox — attention fingerprinting, gradient sensitivity mapping, latent-space (UMAP/PCA) analysis, permutation importance, and partial dependence.

Repository layout

TransMICRO-Net/
├── train.py                 # main training script (3-fold CV + snapshot ensemble + external test)
├── preprocess.py            # extract compression curves from raw Zenodo data into standard tensors
├── downstream_analysis.py   # gradient sensitivity, reconstruction diagnostics, embeddings, cycle stability
├── plotting.py              # all manuscript figures (scatter, violin, attention maps, UMAP, etc.)
├── processed_data/          # preprocessed tensors (curves, labels, metadata)
├── requirements.txt
└── README.md

Data

The raw experimental dataset is publicly available at Zenodo (record 18171138; DOI 10.5281/zenodo.18171138), originally provided by Faber et al. It contains uniaxial compression cyclic tests of OHA-GEL, Alginate, and ADA-GEL hydrogels. preprocess.py converts the raw archive into the tensors stored in processed_data/.

Installation

git clone https://github.com/SHENTongfei/TransMICRO-Net.git
cd TransMICRO-Net
pip install -r requirements.txt

Python 3.9+ recommended; a CUDA GPU is optional (CPU works, slower).

Usage

# 1. (optional) rebuild processed tensors from the raw Zenodo archive
python preprocess.py

# 2. train TransMICRO-Net (3-fold CV + snapshot ensemble + external ADA-GEL test)
python train.py

# 3. downstream interpretability analyses
python downstream_analysis.py

# 4. regenerate manuscript figures
python plotting.py

Outputs are written under models/, logs/, and analysis/ (created automatically).

Key results (manuscript)

• Mean cross-validated R²: 0.824 ± 0.049 (3-fold, 3 indicators).
• External test on unseen ADA-GEL: 0.896 overall; yield point 0.991 and toughness 0.974, top-ranked among all compared models.
• Physics compliance score: 14.95 / 15 on the external test set.

License

For research use. Contact the corresponding authors for details.

Citation

If you use this code or data, please cite:

• Shen, T. et al. TransMICRO-Net: A Quasi-multimodal Multi-task Interpretable Attention-based Deep Learning Framework... (2026).
• Faber, J. et al. Experimental data of OHA-GEL, ADA-GEL and alginate hydrogels for hyperelastic parameter identification. Zenodo, DOI: 10.5281/zenodo.18171138.