Physics-informed neural network for predicting fatigue life of unirradiated and irradiated austenitic and ferritic/martensitic steels under reactor-relevant conditions

This study introduces a Physics-Informed Neural Network (PINN) framework that outperforms traditional machine learning models in predicting the low-cycle fatigue life of both unirradiated and irradiated austenitic and ferritic/martensitic steels by embedding physical constraints to ensure accurate, reliable, and interpretable assessments under reactor-relevant conditions.

The Gist

Imagine you are trying to predict how long a bridge will last. But this isn't a normal bridge; it's a bridge inside a nuclear reactor. It's constantly being hit by invisible bullets (neutrons), baked in extreme heat, and shaken by vibrations. Over time, these forces make the metal brittle and weak, eventually causing it to crack and fail.

Engineers need to know exactly when this will happen to keep the reactor safe. Traditionally, they've had two ways to guess:

1. The "Old School" Math Way: Using complex physics equations. It's accurate but incredibly slow and hard to solve for every new situation.
2. The "Data-Only" Way: Using Artificial Intelligence (AI) to look at past test results and guess the future. It's fast, but if it hasn't seen a specific type of damage before, it might make a wild, dangerous guess.

The Problem: Real-world data on nuclear materials is rare, expensive, and messy. Pure AI often fails when it runs out of data to learn from.

The Solution: The authors of this paper created a "Smart Hybrid" called a Physics-Informed Neural Network (PINN).

Here is a simple breakdown of how it works, using some everyday analogies:

1. The "Student with a Textbook" vs. The "Student with a Cheat Sheet"

• Traditional AI (The Cheat Sheet Student): Imagine a student taking a test who only memorized the answers to the practice questions. If the teacher asks a question slightly different from the practice ones, the student panics and guesses randomly. This is what happens when standard AI tries to predict fatigue life with limited data.
• The PINN (The Student with a Textbook): Now, imagine a student who has memorized the answers but also has the physics textbook open on their desk. They know the fundamental rules of the universe (e.g., "more stress = less life," "higher heat = faster damage"). Even if they haven't seen a specific question before, they can use the rules in the textbook to figure out the logical answer.

2. How the "Physics Rules" Work

The researchers taught their AI the "laws of fatigue" by adding them directly into the computer's brain. They told the AI:

• "If you increase the strain (bending the metal), the life must go down."
• "If you increase the radiation dose, the life must go down."
• "If you increase the temperature, the life must go down."

The AI isn't allowed to break these rules. If it tries to predict that "more heat makes the metal last longer," the system immediately says, "No, that violates physics," and corrects the guess. This keeps the AI honest, even when data is scarce.

3. The Two Types of Steel: The "Fragile Glass" vs. The "Tough Rubber"

The study looked at two main types of steel used in reactors:

• Austenitic Steel (The Fragile Glass): Think of this like a thick glass window. It's strong, but when you hit it with radiation and heat, it gets brittle and cracks easily. The AI found that for this steel, everything matters: how much you bend it, how hot it is, and how much radiation it gets. All three factors work together to break it quickly.
• Ferritic/Martensitic Steel (The Tough Rubber): Think of this like a high-quality rubber band. It's very good at handling radiation (it doesn't get brittle easily). However, it has a "breaking point" with heat. If you keep it cool, it lasts forever. But if you heat it up past a certain point (like 550°C), it suddenly loses its strength and fails fast.

4. Why This Matters

The researchers tested their "Smart Hybrid" AI against other standard AI models (like Random Forests and Gradient Boosting).

• The Result: The PINN was the clear winner. It was more accurate, more stable, and didn't get confused when the data was messy.
• The "Why": Because it wasn't just guessing based on patterns; it was reasoning based on the laws of physics.

The Catch (Limitations)

Even the smartest student needs good study materials. The authors admit that their "textbook" (the dataset) is still a bit incomplete.

• They don't have enough data on the microscopic changes happening inside the metal (like tiny cracks forming at the atomic level).
• Predicting what happens in extreme, never-before-seen conditions is still risky.

The Bottom Line

This paper introduces a new tool that combines the speed of AI with the reliability of physics. It's like giving a supercomputer a rulebook so it can predict when nuclear reactor parts will fail, helping engineers design safer, longer-lasting energy systems without needing to run thousands of expensive, dangerous experiments.

In short: They taught a computer to be a smart engineer, not just a data cruncher, ensuring that when it predicts the future of nuclear materials, it follows the laws of nature.

Technical Summary

1. Problem Statement

Predicting the low-cycle fatigue (LCF) life of structural materials in nuclear reactors is a critical challenge due to the complex interplay of cyclic loading, neutron irradiation, and elevated temperatures.

• Material Degradation: Neutron irradiation creates point defects, dislocation loops, and voids, leading to hardening, embrittlement, and swelling. These mechanisms degrade fatigue resistance differently in austenitic steels (FCC structure, high corrosion resistance) and ferritic/martensitic (F/M) steels (BCC structure, low thermal expansion, high swelling resistance).
• Limitations of Existing Models:
  • Traditional Physics Models: Continuum damage mechanics and viscoplasticity models (e.g., Chaboche) are computationally intensive and require extensive parameter calibration across multiple datasets.
  • Empirical Models: Equations like Manson-Coffin-Basquin rely on costly large-scale testing and often fail to capture complex interactions between alloy composition, temperature, and irradiation dose.
  • Pure Data-Driven ML: Conventional machine learning (ML) models suffer from poor generalization when trained on limited, non-uniform datasets (common in irradiation fatigue testing) and lack physical consistency, leading to unreliable predictions outside the training domain.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework that embeds physical laws directly into the learning process to overcome data scarcity and ensure physical consistency.

A. Dataset and Preprocessing

• Data Source: A curated dataset of 495 strain-controlled LCF data points was compiled from literature, covering both unirradiated and irradiated conditions.
• Materials: Included Austenitic steels (SS304, SS304L, SS310, SS316, SS316L) and F/M steels (HT9, T91, F82H, OPTIFER, EUROFER97).
• Features: Over 50 input features were used, including:
  • Elemental composition (wt.%).
  • Processing parameters (temperature, time, cold-working, annealing, tempering).
  • Irradiation parameters (dose in dpa, temperature, environment: Na, He, Ar, etc.).
  • Testing parameters (strain amplitude, test temperature, strain rate, specimen geometry).
• Preprocessing: Features were normalized (0–1 range). Missing values and infinite values were handled by assigning extreme placeholder values (±10) to distinguish them from actual data gaps without destabilizing training.

B. PINN Architecture and Physics Constraints

The core innovation is the embedding of partial differential equations (PDEs) representing fatigue physics into the loss function.

• Physical Constraints: The model enforces known physical trends via inequality constraints on the partial derivatives of the predicted fatigue life ($N$) with respect to key variables:
  • $\frac{\partial N}{\partial \epsilon} \leq 0$: Fatigue life decreases as strain amplitude ($\epsilon$) increases.
  • $\frac{\partial N}{\partial T} \leq 0$: Fatigue life decreases as test temperature ($T$) increases.
  • $\frac{\partial N}{\partial d} \leq 0$: Fatigue life decreases as irradiation dose ($d$) increases.
  • $\frac{\partial^2 N}{\partial \epsilon^2} \geq 0$: The curvature of the S-N curve is non-negative.
• Loss Function: The total loss ($\mathcal{L}_{total}$) combines:
  1. Data-driven Loss: Huber loss between predicted and experimental fatigue life (robust to outliers).
  2. Physics-informed Loss: A penalty term ($\mathcal{L}_{PDE}$) that penalizes violations of the derivative constraints. This term uses a sigmoid-like transformation $G(x)$ to bound derivatives between 0 and 1.
  3. Regularization: A term to prevent overfitting.
  • The balance is controlled by a hyperparameter $\omega$ (tuned between 0.05 and 0.23).

C. Benchmarking and Validation

• Comparative Models: The PINN was benchmarked against four standard ML models: Random Forest (RF), Gradient Boosting (GB), eXtreme Gradient Boosting (XGB), and a conventional Neural Network (NN).
• Evaluation Metrics: Coefficient of determination ($R^2$) and Mean Squared Error (MSE).
• Validation Strategy:
  • Multiple random train-test splits (80:20 and 70:30) to test robustness.
  • Block-wise cross-validation (5-fold) to ensure generalization across data partitions.
  • SHAP (SHapley Additive exPlanations) analysis for model interpretability.

3. Key Results

A. Predictive Performance

• Superior Accuracy: The PINN achieved the highest test $R^2$ (0.88) and lowest MSE (0.059), outperforming RF, GB, XGB, and NN.
• Generalization & Stability:
  • PINN showed the smallest variance across different data splits ($\sigma \approx 0.020$), whereas conventional models (especially NN) exhibited higher sensitivity to data partitioning.
  • Unlike other models that showed signs of overfitting (high training $R^2$, lower test $R^2$), PINN maintained a tight match between training and testing performance, indicating a superior bias-variance trade-off.
• Cross-Validation: Block-wise cross-validation confirmed stable performance ($R^2$ range 0.76–0.86) across all validation blocks.

B. Interpretability (SHAP Analysis)

• Dominant Features: Strain amplitude, irradiation dose, and test temperature were identified as the top three predictors.
• Physical Consistency: SHAP values confirmed that increases in these three features negatively correlate with fatigue life, aligning with established metallurgical principles.
• Feature Importance: The model correctly prioritized physical variables over processing history, validating that the PINN learned mechanistic relationships rather than spurious correlations.

C. Material-Specific Trends (Univariate & Multivariate Analysis)

Using synthetic data generated by the trained PINN:

• Austenitic Steel (SS316): Exhibited strong nonlinear coupling between strain, dose, and temperature. Fatigue life degraded significantly with increasing irradiation dose and temperature, showing high sensitivity to combined loading.
• F/M Steel (EUROFER97): Demonstrated superior irradiation tolerance.
  • Dose Saturation: Fatigue life remained relatively stable even at high doses (up to 70 dpa), consistent with the "dose-saturation" behavior of F/M steels due to efficient defect recombination in their BCC structure.
  • Temperature Threshold: A sharp decline in fatigue life occurred only when the test temperature exceeded ~550°C (approaching the tempering temperature), indicating microstructural instability (loss of tempered martensite structure) rather than irradiation-induced damage.

4. Significance and Contributions

1. First PINN for Irradiated Fatigue: This study presents the first application of a PINN framework specifically for predicting LCF life in both unirradiated and irradiated austenitic and F/M steels.
2. Data Efficiency: By embedding physical constraints, the model achieves high accuracy with a relatively small dataset (495 points), addressing the scarcity of irradiation fatigue data.
3. Mechanistic Insight: The model successfully distinguishes between the degradation paradigms of FCC (Austenitic) and BCC (F/M) steels, capturing phenomena like dose saturation and temperature-induced microstructural instability without explicit programming of these rules.
4. Reliability: The physics-informed approach prevents the model from generating unphysical predictions (e.g., increasing fatigue life with higher strain), making it a more reliable tool for reactor design and safety assessment.

5. Limitations and Future Outlook

• Data Limitations: The model's reliability is constrained by the non-uniform distribution of the training data and the lack of quantitative microstructural descriptors (e.g., defect density, precipitate size) in the literature.
• Extrapolation Caution: Predictions at extreme irradiation doses (>70 dpa) or very high temperatures should be interpreted with caution as they may fall outside the experimental envelope.
• Future Work: The authors suggest integrating larger datasets enriched with microstructural variables and refining the physics constraints to include more complex mechanisms like irradiation hardening and competing damage modes.

Conclusion

The proposed PINN framework serves as a robust, interpretable, and physically consistent tool for fatigue assessment in nuclear applications. It outperforms traditional data-driven ML models by leveraging physical laws to guide learning, offering a promising pathway for reducing the reliance on expensive post-irradiation testing in the development of advanced nuclear reactors.