🔬 physics

Machine Learning Assisted Creep Deformation Time Prediction and Failure Analysis of 316L Stainless Steel at Elevated Temperatures

This study integrates experimental creep testing, machine learning modeling, and microstructural analysis to accurately predict deformation time and characterize failure mechanisms of 316L stainless steel at elevated temperatures, demonstrating that ensemble algorithms like XGBoost and LightGBM outperform conventional methods in predicting service life while revealing temperature-dependent damage transitions.

Original authors: Arun Kumar, Sanjay Sharma, Biraj Kumar Sahoo, Krishna Guguloth

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arun Kumar, Sanjay Sharma, Biraj Kumar Sahoo, Krishna Guguloth

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

Metals used in power plants, nuclear reactors, and chemical factories face a silent, slow-motion enemy. When these materials are subjected to intense heat and constant pulling forces for years, they do not simply break suddenly; they slowly stretch and deform over time. This process, known as creep, is a gradual wearing down of the material's internal structure that eventually leads to rupture. Engineers must know exactly how long a component can last before it fails, but predicting this lifespan is difficult because the damage accumulates in complex ways that depend on both the temperature and the force applied. Traditional methods try to estimate this life using mathematical rules derived from past experiments, but these rules often struggle to capture the messy, non-linear reality of how metals actually degrade.

In a recent study, researchers at the CSIR-National Metallurgical Laboratory in India and Manav Rachna International Institute of Research and Studies set out to improve this prediction by combining old-school physics with modern computer learning. They focused on 316L stainless steel, a material widely used in high-temperature environments because it resists corrosion and holds its shape well. The team subjected samples of this steel to extreme heat—specifically 675 degrees Celsius and 725 degrees Celsius—while pulling on them with different levels of force. They watched how long it took for the metal to stretch and eventually break, recording every tiny change in shape along the way. Instead of relying solely on traditional formulas, they fed this detailed experimental data into three different types of advanced computer models designed to find patterns in complex information. These models learned to predict the time it took for the metal to deform based on the temperature, the force applied, and the amount of stretching already observed.

The results showed that these computer models could mimic the behavior of the steel with remarkable precision. When the researchers tested the models against data they had not used for training, the predictions were incredibly close to the actual experimental results. One of the models, known as XGBoost, achieved a level of accuracy where its predictions matched the real-world data almost perfectly, explaining more than 99 percent of the variation in how long the steel lasted before failing. Another model, LightGBM, performed even better when the data was shuffled and tested in different groups, suggesting it was very stable. A key finding was that the models became more accurate as the force applied to the steel increased. At lower forces, where the metal lasted for hundreds of hours, the predictions had slightly more uncertainty, likely because the slow, long-term damage processes are harder to pin down. At higher forces, where failure happened more quickly, the models could predict the exact time to failure with errors of less than an hour.

However, the researchers were careful to clarify what these models actually do. They are not magic crystals that can see the future; rather, they are sophisticated tools that reproduce the patterns found in the specific experiments they were trained on. The study explicitly notes that while the computer models are excellent at matching the data they have seen, they do not yet fully understand the physical reasons why the metal fails. To bridge this gap, the team combined their computer predictions with a deep dive into the physical evidence left behind after the steel broke. They examined the fractured surfaces under powerful microscopes and found that the way the metal failed changed depending on the conditions. At the lower temperature, the metal broke in a ductile manner, showing a surface covered in tiny, cup-like depressions formed by microscopic holes merging together. At the higher temperature, the damage was more severe, with visible cracks and cavities forming along the boundaries where the metal grains meet.

Further analysis revealed that this damage was closely linked to tiny particles of chromium and molybdenum that had formed along the grain boundaries during the heat exposure. These particles, which are a type of carbide, created weak spots where cracks could start and grow. The researchers calculated that the energy required for this deformation process was approximately 324 kilojoules per mole, a value that aligns with known mechanisms of atomic diffusion in stainless steel. They also found that the relationship between the force applied and the time to failure followed a specific pattern, with a material constant of 22 providing the best fit for their data. This constant helps engineers create a single curve that predicts how long the steel will last at various temperatures and stresses, a vital tool for safety assessments.

The study concludes that while the computer models are powerful tools for reproducing experimental results, they work best when paired with a physical understanding of the material. The models successfully captured the statistical relationship between temperature, stress, and time, but the real-world failure mechanism was revealed through the microscopic examination of the broken steel. The combination of these two approaches—data-driven prediction and physical failure analysis—offers a more complete picture of how 316L stainless steel degrades. The researchers emphasize that their models are currently limited to the specific range of temperatures and stresses they tested, and they do not yet account for every possible microstructural detail. Nevertheless, this integrated framework provides a robust method for interpreting how high-temperature components degrade, offering a clearer path to ensuring the safety and longevity of critical industrial equipment.

Technical Summary: Machine Learning Assisted Creep Deformation Time Prediction and Failure Analysis of 316L Stainless Steel at Elevated Temperatures

Problem Statement
Accurate prediction of creep deformation and rupture behavior is critical for assessing the structural integrity and remaining service life of metallic components, such as 316L austenitic stainless steel (ASS), operating in high-temperature environments like power generation and nuclear systems. While conventional empirical methods (e.g., Larson-Miller, Manson-Haferd) provide correlations between temperature, stress, and rupture time, they often rely on assumed functional forms and do not explicitly capture the nonlinear relationships between evolving creep strain and deformation time. Furthermore, high statistical accuracy in machine learning (ML) models alone does not guarantee an understanding of the underlying physical mechanisms of creep degradation. There is a need for an integrated framework that combines data-driven prediction with conventional creep kinetics and direct microstructural failure analysis to interpret creep degradation comprehensively.

Methodology
The study employed an integrated experimental and data-driven workflow involving three main components:

Experimental Creep Testing: Constant load creep tests were conducted on 316L ASS at two temperatures (675 °C and 725 °C) under varying applied stresses (175–245 MPa at 675 °C; 120–180 MPa at 725 °C). Tests continued until rupture, with continuous recording of time-dependent creep strain. Post-rupture, specimens underwent microstructural and fractographic analysis using Scanning Electron Microscopy (SEM) and Energy Dispersive Spectroscopy (EDS).
Machine Learning Modeling: Three ensemble ML models—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—were developed. The models used temperature, applied stress, and measured creep strain as input features to predict the experimentally recorded deformation time. The dataset was split 70:30 for training and testing, with five-fold cross-validation performed on the training set.
Conventional Kinetic and Failure Analysis: The experimental data was analyzed using the Norton power-law relationship to determine stress exponents, the Larson-Miller parameter (LMP) for rupture life correlation, and the Arrhenius equation for apparent activation energy. These results were correlated with SEM-EDS observations of grain-boundary damage, carbide precipitation, and fracture morphology.

Key Results

Machine Learning Performance: All three models demonstrated high predictive accuracy in reproducing the experimentally recorded deformation-time response.
- For the combined temperature-stress dataset, XGBoost achieved the highest independent testing coefficient of determination ( $R^2 = 0.9945$ ), with an RMSE of 9.06 h and MAE of 4.60 h.
- LightGBM yielded the highest five-fold cross-validation $R^2$ (0.9956).
- A clear trend was observed where prediction error (RMSE and MAE) decreased as applied stress increased. For instance, at 675 °C, RMSE dropped from ~26 h at 175 MPa to ~2 h at 245 MPa. This was attributed to reduced dispersion in the deformation-time relationship at higher stresses, rather than a fundamental change in model capability.
Conventional Creep Kinetics:
- Stress Exponents ( $n$ ): Calculated as 6.9 at 675 °C and 4.5 at 725 °C, indicating a shift in the stress dependence of the minimum creep rate with temperature.
- Activation Energy ( $Q_c$ ): The apparent creep activation energy was determined to be approximately 324 kJ mol⁻¹.
- Larson-Miller Parameter: The strongest correlation ( $R^2 = 0.990$ ) was achieved with a material constant $C = 22$ .
- Threshold Stress: Apparent threshold stresses were identified as ~172.2 MPa at 675 °C and ~110 MPa at 725 °C.
Microstructural and Fractographic Analysis:
- Fracture Morphology: At 675 °C/175 MPa, the fracture surface exhibited predominantly dimpled morphology indicative of microvoid coalescence. At 725 °C/120 MPa, the morphology shifted to more pronounced cavity- and crack-associated damage, suggesting increased grain-boundary (GB) involvement.
- Microstructural Damage: SEM-EDS analysis revealed Cr and Mo-enriched regions adjacent to damaged grain boundaries, consistent with the formation of $M_{23}C_6$ type carbides. These precipitates were associated with localized GB damage, cavity nucleation, and crack propagation.

Significance and Claims
The authors claim that the study establishes an integrated experimental–data-driven framework for interpreting the creep degradation and failure of 316L stainless steel. The significance of this work lies in its dual approach:

Quantitative Prediction: It demonstrates that ensemble ML models can accurately reproduce complex, nonlinear creep deformation-time relationships within the investigated domain.
Physical Interpretation: It explicitly links these statistical predictions to conventional kinetic parameters and direct physical evidence of failure mechanisms (carbide precipitation, GB damage, and cavity growth).

The paper modestly defines the scope and limitations of its ML framework:

The models are designed to reproduce experimentally recorded deformation times based on specific input variables (temperature, stress, strain) rather than serving as unrestricted prospective predictors of unknown rupture lives.
The high $R^2$ values reflect strong consistency within the available dataset but do not constitute definitive proof of independent specimen-level generalization, as the cross-validation was performed at the data-point level rather than the creep-curve level.
The models do not explicitly include microstructural descriptors (e.g., grain size, precipitate population) as inputs, limiting their application to the specific experimental domain used for training.

In conclusion, the study posits that combining ML prediction with conventional creep kinetics and post-creep damage characterization provides a more robust and physically informed method for analyzing creep behavior than either approach alone.

More like this