Knowledge-infused Probabilistic Modelling for Decision Support in Metal Additive Manufacturing
- Halid Can YILDIRIM
- Dec 27, 2025
- 3 min read
A hybrid Bayesian framework integrating physics-informed yield strength corrections with neural networks for fatigue life prediction of additively manufactured metals.
Overview:
This repository implements a novel probabilistic framework for fatigue life prediction in additively manufactured (AM) metals, combining Bayesian Neural Networks with a physics-informed yield strength correction procedure. The methodology addresses critical limitations of conventional deterministic approaches by quantifying both aleatoric (material variability) and epistemic (model uncertainty) uncertainty components, providing statistically rigorous confidence bounds essential for risk-informed design.
By integrating material-specific yield strength normalisations with mean stress corrections, this framework enables accurate cross-material fatigue life predictions while maintaining physical interpretability. The Bayesian architecture systematically captures the inherent variability in AM processes, offering significant improvements over conventional methods while supporting reliable decision-making for safety-critical applications.
Key Challenges Addressed:
Material heterogeneity: Process-induced variations in yield strength across different AM alloys (Ti-6Al-4V, AlSi10Mg, 316L, IN718)
Uncertainty quantification: Comprehensive modeling of both aleatoric (data-driven) and epistemic (model-driven) uncertainty sources
Cross-material generalizability: Framework extensibility across multiple alloy systems with varying fatigue characteristics
Limited experimental data: Robust performance on small to moderate-sized datasets typical in fatigue testing
Load ratio variability: Accurate predictions across different loading conditions (R = 0.0 to R = -1.0)
Key Features:
Hybrid Bayesian Architecture: Unified framework combining Bayesian Neural Networks with physics-informed corrections
Yield Strength Correction: Novel procedure accounting for process-induced material property variations
Dual Uncertainty Quantification: Explicit separation and modeling of aleatoric and epistemic uncertainties
Four Model Configurations:
Case 1: Deterministic prediction (baseline)
Case 2: Aleatoric uncertainty only (data variability)
Case 3: Epistemic uncertainty only (model uncertainty)
Case 4: Combined aleatoric and epistemic uncertainty (complete probabilistic)
Physical Consistency: Integration of SWT mean stress corrections with material-specific yield strength adjustments
Scalable Implementation: Efficient training and inference suitable for industrial applications Applications:
The framework supports diverse applications across multiple domains:
Industrial Sectors:
Aerospace: Fatigue life prediction for AM aircraft components
Automotive: Reliability assessment of additively manufactured automotive parts
Biomedical: Design optimization for AM implants and medical devices
Energy: Life prediction for AM components in power generation equipment
Research & Development:
Material Development: Accelerated testing and characterization of new AM alloys
Process Optimization: Identifying optimal processing parameters for fatigue resistance
Digital Twins: Integration with digital twin systems for in-service life prediction
Certification Support: Data-driven approach for AM component qualification
Education & Benchmarking:
Teaching Tool: Demonstrating probabilistic methods in materials engineering
Benchmark Dataset: Standardized framework for comparing fatigue prediction methods
Methodology Development: Foundation for extending to other material systems Validation:
The framework has been extensively validated against experimental data:
Dataset Size: 10,501 high-quality fatigue data points across four major AM alloys
Materials: Ti-6Al-4V (5,392 points), AlSi10Mg (1,762 points), 316L (1,599 points), IN718 (1,748 points)
Loading Conditions: Load ratios from R = 0.0 to fully reversed R = -1.0
Performance Metrics:
RMSE reduction: 28-42% compared to conventional methods
R² improvement: 0.68-0.74 (deterministic) to 0.86-0.88 (combined uncertainty)
Prediction coverage: 92% of experimental data within 95% confidence intervals Contributions:
Contributions to model enhancement, experimental validation, or industrial applications are welcome. Please contact the authors for collaboration opportunities.
Citation:
If this work supports your research, please cite the original paper: @article{YILDIRIM2025115212,
title = {Knowledge-Infused Probabilistic Modelling for Decision Support in Metal Additive Manufacturing},
journal = {Knowledge-Based Systems},
pages = {115212},
year = {2025},
issn = {0950-7051},
doi = {https://doi.org/10.1016/j.knosys.2025.115212},
url = {https://www.sciencedirect.com/science/article/pii/S0950705125022464},
author = {Halid Can Yıldırım},
keywords = {Additive Manufacturing (AM), Fatigue Life Prediction, Probabilistic Modelling, Bayesian Inference},
abstract = {This study presents a probabilistic framework for fatigue life prediction in additively manufactured (AM) metals by integrating Bayesian neural networks with a novel physics-informed yield strength correction. The framework analyses 10,501 fatigue data points across Ti-6Al-4V, AlSi10Mg, 316L, and IN718 alloys under varying loading conditions (from R=0.0 to R=−1), addressing critical limitations of conventional deterministic methods. The yield strength correction accounts for process-induced material variations, while the Bayesian approach quantifies both aleatoric (material variability) and epistemic (model uncertainty) uncertainty components. Results demonstrate significant improvements over conventional predictions, with fatigue strength corrections up to 7 × at 2 million cycles. The hybrid methodology reduces prediction errors by 28-42% while providing statistically rigorous confidence bounds that capture 92% of experimental data within 95% prediction intervals. This integrated approach enables more reliable AM component design through its combination of physical interpretability and comprehensive uncertainty quantification, offering particular value for safety-critical applications where traditional methods prove overly conservative.}
} Basic Usage:
# Install dependencies
pip install -r requirements.txt
# Update configuration paths in Main_script.py
# Set file_path to your Excel data file
# Set base_output_directory for results
# Run complete framework
python main_script.py
Comments