⚡ electrical engineering

Active Learning-Driven Surrogate Modeling for Fatigue LifePrediction of LPBF Hastelloy-X Using Defect-Informed FiniteElement Analysis

This paper presents a hybrid framework that integrates defect-informed finite element analysis with active learning-driven Gaussian process surrogate modeling to accurately and efficiently predict the fatigue life of LPBF Hastelloy-X under various conditions, achieving high fidelity with over 95% reduction in experimental data requirements.

Original authors: Chandrashekhar Pilgar, G Gangaraju

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chandrashekhar Pilgar, G Gangaraju

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how long a specific type of metal part will last before it cracks under stress. This metal, called Hastelloy-X, is often used in jet engines and power plants because it can handle extreme heat. However, when this metal is made using a high-tech 3D printing method called Laser Powder Bed Fusion (LPBF), it comes out with tiny, invisible "scars" and "bumps" that make it behave differently than traditional metal.

The researchers in this paper faced a massive problem: To figure out exactly how long these 3D-printed parts last, you usually have to build hundreds of them and break them in a machine over and over again. This is incredibly expensive, slow, and wasteful.

Here is how they solved it, explained simply:

1. The "Scars" of 3D Printing

Think of 3D-printed metal like a cake baked in a very specific way.

Surface Roughness: Instead of being smooth like a machine-milled part, the surface looks like a staircase or a bumpy road. The direction you print the part (upright, flat, or on a slant) changes how bumpy it is.
Porosity (Tiny Holes): Inside the metal, there are tiny air bubbles (like bubbles in a sponge).
The Problem: These bumps and bubbles act like stress concentrators. If you pull on the metal, the cracks start at the deepest valleys of the bumps or the edges of the bubbles.

2. The Old Way vs. The New Way

The Old Way (The "Brute Force" Method): To predict how long the part lasts, engineers used to test hundreds of samples at different temperatures and angles. They would build a massive database of broken parts. It's like trying to learn a new language by memorizing every single word in the dictionary before you can speak a sentence.
The New Way (The "Smart Detective" Method): The authors created a hybrid system that acts like a smart detective. They combined three tools:
1. A Physics Formula (UDSL): A mathematical rulebook that explains why the metal breaks based on heat, strain, and bumps.
2. A Virtual Simulator (FEA): A computer program that acts like a wind tunnel, simulating how stress flows around the bumps and holes without actually breaking a physical part.
3. A Smart Learner (Active Learning): An AI that decides which experiments to run next. Instead of guessing randomly, it asks, "Where am I most confused?" and only runs a test there to learn the most.

3. The Big Discovery: The "Sub-Additive" Effect

One of the coolest findings is about how the bumps and holes interact.

The Intuition: You might think that if you have a bumpy surface and a hole underneath it, the damage is just "Bump Damage + Hole Damage."
The Reality: The researchers found that the damage is actually less than the sum of the two. They call this "sub-additive."
The Analogy: Imagine two people shouting at you. If they shout from the exact same spot, it's deafening. But if one is shouting from a hill and the other is in a valley, their voices don't amplify each other as much because the sound waves interfere. Similarly, the stress fields from the surface bumps and the internal holes partially cancel each other out rather than stacking up to destroy the part.

4. The Result: A Massive Time Saver

The team tested their new system against a massive dataset of 440 actual fatigue tests (breaking parts).

The Physics Model: They calibrated their "Rulebook" using only 19 data points. It predicted the results of the other 421 tests with 98% accuracy.
The AI Learner: They used the "Smart Detective" AI. By only testing 20 carefully chosen samples (instead of 440), the AI learned the pattern just as well as if it had seen them all.
The Savings: This approach reduced the amount of physical testing needed by over 95%.

5. Why This Matters (According to the Paper)

The paper claims this is a "hybrid framework" that is perfect for safety-critical parts (like those in jet engines).

It's Fast: You don't need to wait years to get data.
It's Honest: The AI doesn't just guess a number; it tells you how confident it is in that guess (e.g., "I'm 95% sure this part will last this long").
It's Adaptable: It works for different temperatures (from room temp to 750°C) and different printing angles.

In a nutshell: The researchers built a "smart shortcut" that uses physics and a little bit of AI to predict how long 3D-printed metal parts will last. They proved that you don't need to break hundreds of parts to know the answer; you only need to break a few dozen, provided you have the right mathematical map and a smart guide to tell you where to look next.

Technical Summary: Active Learning-Driven Surrogate Modeling for Fatigue Life Prediction of LPBF Hastelloy-X

Problem Statement
Additively manufactured Hastelloy-X produced via Laser Powder Bed Fusion (LPBF) exhibits fatigue behavior highly sensitive to process-induced defects, including surface roughness, subsurface porosity, and build orientation. Existing life-prediction approaches typically treat these factors independently, rely on extensive experimental datasets, and fail to account for the complex, nonlinear interactions between defects and thermo-mechanical loading conditions. Furthermore, purely data-driven machine learning models often suffer from poor extrapolation capabilities and a lack of uncertainty quantification, while classical empirical models (e.g., Coffin-Manson-Basquin) require separate parameter sets for every combination of material, orientation, and temperature, rendering them inefficient for the vast design space of LPBF components. The experimental burden for characterizing fatigue across multiple strain amplitudes, temperatures (up to 750°C), and orientations is prohibitively high, often requiring hundreds of specimens.

Methodology
The authors propose a hybrid, defect-informed framework that integrates physics-based modeling, high-fidelity Finite Element Analysis (FEA), and data-efficient Gaussian Process (GP) active learning. The methodology consists of three primary components:

Unified Defect-Sensitive Strain-Life (UDSL) Model:
A physics-informed analytical model was developed to predict strain-controlled fatigue life ( $N_f$ ) across a wide design space. Unlike arbitrary polynomial regressions, the UDSL model derives terms from specific physical mechanisms:
- Strain-Life Power Law: Generalized Basquin/Coffin-Manson relations.
- Roughness as a Distributed Notch: Modeled using a logarithmic term based on Peterson's notch sensitivity, linking arithmetic mean roughness ( $R_a$ ) to effective stress concentration.
- Orientation Geometry: A term accounting for the "staircase effect" dependent on build angle ( $\theta$ ).
- Homologous Temperature Scaling: Normalizes temperature against the melting point ( $T_m = 1628$ K) to capture thermally activated deformation.
- Strain-Temperature Interaction: Cross-terms capturing the shift from elastic to plastic dominance at elevated temperatures.
  The model was calibrated using only 19 data points from a single condition (27°C, horizontal orientation) and validated against a comprehensive 440-point dataset.
Defect-Informed Finite Element Analysis (FEA):
High-fidelity 3D simulations in Abaqus were conducted to quantify stress concentration factors ( $K_t$ ) and defect interactions:
- Surface Roughness: Realistic LPBF surfaces were generated via Fourier synthesis to calculate $K_t$ , validating the roughness coefficients in the UDSL model.
- Porosity and Void Growth: Gurson-Tvergaard-Needleman (GTN) porous plasticity models demonstrated significant void growth (7.3x amplification) under plastic strain, justifying the inclusion of porosity-strain interaction terms.
- Combined Defect Interaction: Simulations of subsurface pores located beneath rough surface valleys revealed a sub-additive coupling effect. The combined stress concentration was lower than the sum of individual effects, leading to a generalized interaction formulation with a calibrated interaction coefficient ( $\rho = 0.329$ ).
Gaussian Process (GP) Surrogate with Active Learning:
To address data scarcity and provide uncertainty quantification, a GP surrogate was trained on physics-informed features derived from the UDSL model and FEA results (e.g., homologous temperature, effective strain, stress concentration factors). An active learning (AL) strategy was employed, iteratively selecting the most informative experimental points based on maximum predictive variance. This approach aimed to minimize the number of required fatigue tests while maximizing model accuracy.

Key Results

Experimental Findings: Fatigue life decreased monotonically with temperature, dropping by nearly two orders of magnitude between 27°C and 750°C. A consistent orientation-dependent ranking was observed: Horizontal > 45° > Vertical. Surface roughness was identified as the dominant factor governing this anisotropy, whereas bulk porosity showed no systematic correlation with fatigue life under the tested surface conditions.
UDSL Model Performance: The 8-parameter UDSL model achieved an $R^2$ of 0.984 and a Root Mean Square Error (RMSE) of 0.131 (in $\log_{10} N_f$ ) on the full 440-point dataset. It predicted 97.7% of the data within a factor of two of the experimental life. Crucially, this accuracy was achieved using only 19 calibration points, representing a >95% reduction in experimental data requirements compared to conventional calibration methods.
FEA Insights: The FEA confirmed that the interaction between surface roughness and subsurface porosity is sub-additive. The combined stress concentration ( $K_t \approx 2.85$ ) was significantly lower than the multiplicative prediction, providing a less conservative and more accurate basis for design.
Active Learning Efficiency: The GP surrogate, trained on physics-informed features, reached the same predictive accuracy ( $R^2 = 0.984$ ) as the UDSL model using only 20 actively selected training points. This represents a 2.5x improvement in data efficiency compared to random sampling, which required approximately 50 points to achieve similar accuracy.
Uncertainty Quantification: The GP framework provided well-calibrated uncertainty estimates, with nominal 95% confidence intervals capturing 94.2% of validation data. This allows for risk-informed decision-making and targeted experimental planning.

Significance and Claims
The paper claims to address three critical gaps in the current state of the art:

Coupled Defect Modeling: It is the first model to simultaneously and quantitatively account for the coupled effects of surface roughness, subsurface porosity, build orientation, and temperature on LPBF Hastelloy-X fatigue life.
Defect Interaction Characterization: It provides the first high-fidelity FEA characterization of the stress field interaction between roughness and porosity, establishing a sub-additive interaction rule rather than assuming additive or independent effects.
Integrated Framework: It demonstrates a practical framework that integrates physics-based modeling, defect-informed simulation, and data-efficient active learning.

The authors assert that this approach reduces the experimental burden by more than 95% while maintaining high predictive fidelity and providing calibrated uncertainty. The integration of physical interpretability (via UDSL), defect sensitivity (via FEA), and uncertainty quantification (via GP) offers a scalable pathway for the qualification of LPBF components in safety-critical, high-temperature applications. The study concludes that while the UDSL model excels in extrapolation and interpretability, the GP+AL framework offers superior adaptability for probabilistic assessment and adaptive test planning, with the two approaches being complementary rather than competing.