AI APPROACHES FOR ANALYZING THE IMPACT OF ADDITIVE MANUFACTURING PROCESSES ON FATIGUE BEHAVIOR��INTERNSHIP DEFENSE

JESSIE KREINSEN

​

M1 NUCLEAR ENGINEERING

(PHYSICS TRACK)

​

ARTS ET MÉTIERS, I2M BORDEAUX

INSTITUT POLYTECHNIQUE DE PARIS | ENSTA PARIS

�AI FOR FATIGUE LIFE PREDICTION

Objectives

• Develop a probabilistic framework linking additive manufacturing (AM) defects to fatigue life.
• Validate the framework using fracture mechanics and synthetic data.
• Develop a physics-informed neural network (PINN) for predicting fatigue life.
• Evaluate the minimum data required for reliable fatigue-life prediction.

Challenges

• Stochastic defect populations control fatigue.
• Experimental fatigue datasets are limited and expensive to obtain.
• Machine learning must remain consistent with fracture mechanics.

OVERALL WORKFLOW

The two workflow branches:

• The synthetic framework (left) links AM defects to fatigue through a probabilistic framework and verifies the physics.
• The experimental branch (right) adapts the framework to real fatigue data and trains the PINN.

SYNTHETIC MODEL WORKFLOW

BUILDING THE SYNTHETIC DEFECT POPULATION

CRITICAL DEFECT:�GENERALIZED EXTREME VALUE (GEV) MODEL

PHYSICS VALIDATION:�PARIS LAW FATIGUE MODEL

PINN ARCHITECTURE & TRAINING SETUP

GENERAL PINN FRAMEWORK FROM LIAO ET AL. (2025).

FROM SYNTHETIC TO REAL DATA:�INITIAL EXPERIMENTAL VALIDATION

FROM SYNTHETIC TO REAL DATA:�SCALING UP

• Zhang & Xu (2023) compiles a much larger, multi-source L-PBF fatigue database
• Filter for a consistent Ti-6Al-4V, L-PBF subset
• R = 0.1, fatigue temperature = 25°C, environment = “air”, build direction = 90°
• Columns standardized: stress, defect position, ΔK, process parameters, fatigue life
• Size of dataset:
• Before cleaning: 111,009 specimens
• After cleaning: 10,886 specimens
• Re-derive the physics directly from the database
• VED-defect density trend directly from database
• Defects generated, largest defect for each specimen identified

​

FEATURE ENGINEERING:�FOUR MODEL VARIANTS

• Progressively deeper inputs test what actually drives fatigue-life scatter:
• Model A: physics-only 🡪 too little info
• log(defect size), log(stress amplitude), Y
• Model B: add process 🡪 too little info
• + log(VED)
• Model C: add loading 🡪 chosen for further analysis for its relevancy and level of detail
• + log(Δσ), log(frequency)
• Model D: add material / mechanical 🡪 too much info 🡪 risk greater overfitting
• + stress concentration factor (K_t ), Young’s modulus (E), yield strength, UTS, elongation

PINN TRAINING RESULTS:�LOSS HISTORY

• Data loss decreases steadily, indicating improved agreement with the experimental data.
• Physics loss decreases initially but then plateaus, remaining the dominant contribution to the total loss.
• The network learns fatigue-life trends beyond the analytical Paris-law baseline.
• The persistent physics loss suggests that Paris law alone does not fully explain the variability in the experimental database.

RESULTS: �PREDICTIVE PERFORMANCE

• Training and test predictions show similar behavior.
• Predictions closely follow the 1:1 trend across the fatigue-life range.
• Validation RMSE reaches a minimum before gradually increasing, indicating mild overfitting.
• The PINN captures the overall defect–fatigue relationship and generalizes reasonably well.

RESULTS:�DATA AVAILABILITY STUDY

• Model retrained on 1–70% of the database across 20 random train/test splits per fraction.
• Error and run-to-run variability both fall sharply between 1% and 10% of the data (1 to 16 samples).
• Model performance improves rapidly with the first 10–20% of the data (16-33 samples), after which returns diminish.
• 20% 🡪 40% only reduces median log(RMSE) from 1.49 to 1.37
• Most of the benefit of the large database is captured with a modest training subset.

CONCLUSIONS

THANK YOU FOR LISTENING

SOURCES

Bittner, F., Müller, B., and Thielsch, J. (2022). Efficient LPBF-process development by design of experiments. Technical report, Fraunhofer Institute for Machine Tools and Forming Technology (IWU), Dresden, Germany.

Liao, e. a. (2025). A physics-informed neural network method for identifying parameters and predicting remaining life of fatigue crack growth. International Journal of Fatigue, 191:108678.

Murakami, Y. (2019). Metal Fatigue: Effects of Small Defects and Nonmetallic Inclusions. Academic Press, Oxford, United Kingdom, second edition edition.

Paris, P. and Erdogan, F. (1963). A critical analysis of crack propagation laws. Journal of Basic Engineering, 85(4):528–533.

Shimatani, Y., Shiozawa, K., Nakada, T., and Yoshimoto, T. (2010). Effect of surface residual stress and inclusion size on fatigue failure mode of matrix HSS in very high cycle regime. Procedia Engineering, 2:873–882.

Wang, H. et al. (2022). Fatigue performance at ultra-low porosity of Ti6Al4V produced by laser powder bed fusion after post heat treatment. SSRN Electronic Journal.

Zhang, Z. and Xu, Z. (2023). Fatigue database of additively manufactured alloys. Scientific Data, 10:249.

Zhou, S. et al. (2025). A general physics-informed neural network framework for fatigue life prediction of metallic materials. Engineering Fracture Mechanics, 322:111136.

APPENDIX:�SYNTHETIC DATA INPUTS

Material & Geometry Parameters

Loading Conditions & VED

APPENDIX:�VED–RELATIVE DENSITY CORRELATION

For relative densities and optimal VED: Bittner, F. et al. (2022) and Park, H. et al. (2024).

For expected relative density vs. VED curve: Park, H. et al. (2024).

APPENDIX:�DISTRIBUTION FITTING: IDENTIFYING THE DEFECT SIZE PDF

• Candidate distributions:
• Lognormal
• Weibull
• Gumbel
• Selection criterion:
• Maximum Likelihood Estimation (MLE)
• Minimum Kolmogorov-Smirnov statistic
• Result:
• Best fit: lognormal, Parameters: mu_ln = -9.904, sigma_ln = 0.3579
• Confirms assumption for defect population
• Best-fit distribution used to represent defect population

​

APPENDIX: �WORKFLOW OUTPUTS

FATIGUE LIFE AND STRESS INTENSITY FACTOR RELATIONSHIP IS CALCULATED WITH THE SHIOZAWA APPROXIMATION.

APPENDIX: �RELIABILITY ANALYSIS: PROBABILITY OF SURVIVAL

APPENDIX:�PSN SENSITIVITY, AND VARIABILITY CURVES