This repository presents a comprehensive workflow for predicting ductile-to-brittle transition temperature (DBTT) and upper shelf energy (USE) in steels using an integrated attention-based deep learning framework. The approach combines physics-informed feature engineering, attention-enhanced multilayer perceptrons (MLP), and Evidential Deep Learning (EDL) to achieve accurate and uncertainty-aware predictions, facilitating reliable evaluation of structural integrity in critical applications such as pipelines, pressure vessels, and cryogenic systems.
Dataset: 4,105 experimental samples comprising 167 descriptors, including Magpie-derived features, compositional variables, impact energy, and test temperature.
Descriptors: Physics-informed and statistically curated features embedding compositional, microstructural, and thermomechanical information.
Models:
- Attention-Enhanced MLP: Learns nonlinear feature–property relationships with adaptive weighting through an attention mechanism.
- Evidential Deep Learning (EDL): Extends the MLP to jointly predict property values and associated aleatoric and epistemic uncertainties.
- Baselines: Ensemble regressors—Random Forest, Extra Trees, and XGBoost—for comparative evaluation.
Evaluation Metrics: MAE, RMSE, R², and Negative Log-Likelihood (NLL) to assess prediction accuracy and uncertainty calibration.
├── dataset/ # Dataset files
├── MLP/ # Scripts for attention-based MLP and EDL training
├── mlp_models/ # Saved MLP and EDL model checkpoints
├── Traditional ML/ # Random Forest, XGBoost, and Extra Trees training scripts
└── README.md
- Curated dataset containing 4,105 samples with 167 descriptors.
- Integration of compositional, impact energy, and temperature data.
- Standardization and outlier filtering to ensure model robustness.
- Physics-informed and Magpie-derived features capturing composition–structure–property relationships.
- Redundancy reduction via correlation and variance filtering.
- Domain-guided feature selection to preserve interpretability.
- Attention-Enhanced MLP:
- Layers: Dense + Attention + Dropout
- Optimizer: Adam with learning rate scheduling
- Hyperparameter tuning using Optuna
- Evidential Deep Learning (EDL):
- Output: Normal–Inverse–Gamma (μ, ν, α, β) parameters
- Captures both aleatoric and epistemic uncertainty
- Penalized loss to balance accuracy and uncertainty calibration
- Baselines:
- Random Forest, Extra Trees, and XGBoost regressors trained on identical features for fair comparison.
- Metrics: MAE, RMSE, R², and NLL.
- EDL demonstrated improved uncertainty calibration and overall predictive reliability.
- Integration of attention mechanisms within MLP for selective feature relevance learning.
- Application of Evidential Deep Learning for simultaneous prediction and uncertainty estimation.
- Development of a physics-informed, data-driven framework for fracture behavior modeling.
- Establishes a reliable and interpretable predictive pipeline for materials design and screening.