PITCH: Physics-Informed Transformer for Hysteretic Response Prediction
A physics-informed transformer with token classification for hysteretic response prediction and uncertainty quantification of RC columns.
- NRMSE
- 0.055
- Test median (n=38)
- Correlation
- 0.986
- Test set
- Peak error
- 6.7%
- Shear force
- Parameters
- 247K
- Compact model
Overview
This project develops a novel deep learning architecture, a Tokenized Bidirectional Transformer with FiLM Conditioning, that predicts the complete cyclic (hysteretic) shear force-drift response of reinforced concrete columns.
Novel contributions
- 01
Soft-decode bridge
Differentiable map from token logits to continuous shear values via temperature-scaled expected value.
- 02
FiLM conditioning
Per-layer Feature-wise Linear Modulation from metadata embedding.
- 03
Native uncertainty quantification
Softmax distribution over 256 bins yields calibration-free uncertainty.
Model architecture
Sample results
- 2026PublishedView article
Physics-Informed Transformer with Token Classification for Hysteretic Response Prediction and Uncertainty Quantification of Reinforced Concrete Columns
Wakjira, T.G., Goshu, H.L.
Applied Soft Computing, 202, 115955
Interested in this research?
For code access, collaboration or questions about PITCH, contact the principal investigator, Dr. Tadesse Wakjira.