🔬 physics

Constitutive Modeling Method for AM 316 Based on Johnson-Cook Guided PCA-MLP Residual Learning

This paper proposes a physics-informed data-driven constitutive modeling approach that combines the Johnson–Cook model with PCA-MLP residual learning to achieve high-accuracy and stable predictions of dynamic flow stress for additively manufactured 316L stainless steel across various temperatures and strain rates.

Original authors: zequn lv, an he, jining li, yihang fan, hang zhao, yongshuo xu, wenchang fu, yingbo wang, xiangyong li

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: zequn lv, an he, jining li, yihang fan, hang zhao, yongshuo xu, wenchang fu, yingbo wang, xiangyong li

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 a material that is strong enough to hold up a jet engine, resistant enough to survive the corrosive depths of the ocean, and flexible enough to be shaped into complex, lightweight parts that traditional manufacturing simply cannot create. This is 316L stainless steel made through additive manufacturing, a process where metal is built layer by layer, much like a 3D printer. While this technology offers incredible freedom in design, the way the metal behaves when it is hit hard or heated quickly is surprisingly different from the same steel made in a factory. When this material is subjected to the extreme conditions of high-speed impact or intense heat, its internal structure shifts in complex ways, causing it to harden, soften, and flow in patterns that are difficult to predict. Engineers need to know exactly how this material will react to ensure that the aircraft, nuclear components, or defense systems built from it will not fail when it matters most.

For decades, scientists have relied on mathematical formulas to guess how metals will behave under such stress. One of the most famous of these formulas, known as the Johnson-Cook model, treats the effects of stretching, speed, and heat as separate factors that simply multiply together. However, for this specific type of 3D-printed steel, the reality is far more tangled. The heat generated by the metal deforming, the speed at which it is stretched, and the microscopic changes inside the metal all influence each other in a way that simple formulas cannot capture. On the other hand, modern computer programs that learn from data, known as machine learning, can find complex patterns, but they often struggle when they do not have enough examples to study. They might memorize the data they are given but fail to make accurate guesses about new situations, such as temperatures they have never seen before.

To solve this puzzle, a team of researchers from universities in China developed a new approach that combines the best of both worlds. They started with the traditional Johnson-Cook formula to provide a solid, physical foundation—a rough sketch of how the metal should behave. Then, instead of trying to force a computer to learn the entire complex story from scratch, they asked the computer to focus only on the mistakes the traditional formula made. They calculated the difference between what the formula predicted and what actually happened in their experiments, creating a "residual" or a leftover error curve. This leftover curve represents the complex, hidden interactions that the simple formula missed.

The researchers then used a clever mathematical technique to simplify these leftover curves. Imagine taking a long, wiggly line and finding just a few key shapes that, when added together, recreate the whole line. This process, called principal component analysis, allowed them to turn a massive amount of data into a small, manageable set of numbers. They then trained a neural network, a type of computer program designed to recognize patterns, to predict these few key numbers based on the temperature and speed of the impact. Once the computer predicted the numbers, the researchers used them to reconstruct the full, corrected curve and added it back to the original formula. The result was a hybrid model that kept the physical reliability of the old formula while using data to fix its specific errors.

The team tested this new method on 316L stainless steel that had been printed using a laser-directed energy deposition process. They subjected the metal to high-speed compression tests at various temperatures, ranging from 150 degrees Celsius to 620 degrees Celsius, and at strain rates as high as 5,000 per second. To truly test the model's ability to generalize, they deliberately left out two specific temperatures, 272 degrees Celsius and 450 degrees Celsius, from the training data. They wanted to see if the model could accurately predict how the metal would behave at these unseen temperatures without ever having seen an example of them before.

The results were striking. When the model was asked to predict the behavior of the metal at 272 degrees Celsius, it achieved an average accuracy score of 0.9306, with an average error of just 1.97 percent. At 450 degrees Celsius, the performance was even better, with an accuracy score of 0.9538 and an average error of only 1.63 percent. In contrast, the traditional formula alone performed poorly at these unseen temperatures, often producing errors that were more than ten times larger. Even the pure computer-learning models, which tried to learn the entire behavior without the help of the physical formula, struggled significantly, sometimes failing to capture the correct trend entirely. The new hybrid method, however, maintained high accuracy and stability across all conditions, successfully reconstructing the complex flow of stress and strain even for temperatures it had never encountered.

This work demonstrates that by guiding a powerful data-driven tool with a trusted physical foundation, scientists can create models that are both accurate and reliable, even when data is scarce. The researchers found that their method does not just memorize past experiments; it learns the underlying rules of how the material deviates from simple expectations. This approach offers a practical path forward for engineers who need to simulate and design with advanced materials, ensuring that the next generation of aerospace and energy systems can be built with confidence. The study confirms that for complex materials like 3D-printed steel, the future of prediction lies not in choosing between physics and data, but in weaving them together so that each strengthens the other.

Technical Summary: Constitutive Modeling Method for AM 316 Based on Johnson-Cook Guided PCA-MLP Residual Learning

Problem Statement
Accurate prediction of the dynamic flow stress for additively manufactured (AM) 316L stainless steel under high temperatures and high strain rates is essential for service safety evaluation and process optimization. However, conventional empirical constitutive models, such as the Johnson-Cook (JC) model, often fail to capture the complex thermo-mechanical coupling and nonlinear interactions inherent in AM materials due to their microstructural heterogeneity (e.g., columnar grains, cellular substructures). Conversely, purely data-driven models (e.g., Multilayer Perceptrons or MLPs) struggle with generalization when trained on small datasets, often lacking physical consistency and exhibiting instability in unseen conditions. The challenge lies in developing a modeling strategy that retains the physical interpretability of empirical models while leveraging the nonlinear fitting capabilities of machine learning to correct systematic deviations.

Methodology
The authors propose a hybrid framework termed JC-PCA-MLP, which integrates a physical baseline with data-driven residual learning. The methodology proceeds through the following steps:

Physical Baseline (Johnson-Cook): The classical JC model is employed to generate a baseline prediction of flow stress. This model captures the dominant physical trends of strain hardening, strain rate strengthening, and thermal softening but is acknowledged to be insufficient for the complex coupling effects in AM 316L.
Residual Construction: The difference between the experimentally measured flow stress and the JC-predicted stress is defined as the "residual curve." This shifts the learning task from predicting the entire stress-strain curve to predicting the deviation from the physical baseline.
Dimensionality Reduction (PCA): To address the high dimensionality of full stress-strain curves and the limitations of small sample sizes, Principal Component Analysis (PCA) is applied to the residual curves. This reduces the residual data to a small set of principal component coefficients, preserving the dominant variation patterns of the residuals.
Machine Learning (MLP): A Multilayer Perceptron (MLP) is trained to map input features (temperature, strain rate, and their interaction terms) to the PCA coefficients of the residuals. The input features include logarithmic strain rates and quadratic temperature terms to better capture nonlinear variations.
Curve Reconstruction: The final predicted stress-strain curve is reconstructed by superimposing the MLP-predicted residual (reconstructed via PCA inverse transformation) onto the JC baseline.
Training Strategy: The model utilizes a composite loss function that simultaneously minimizes the error in predicting PCA coefficients and the reconstruction error of the final stress-strain curves.

Experimental Validation
The study utilized dynamic compression data from Split Hopkinson Pressure Bar (SHPB) tests on AM 316L stainless steel fabricated via Laser Directed Energy Deposition (L-DED). The dataset covered temperatures of 150°C, 272°C, 450°C, and 620°C, and strain rates of 3000 s⁻¹, 4000 s⁻¹, and 5000 s⁻¹.
To rigorously test generalization, a leave-one-intermediate-temperature-out validation strategy was employed. The model was trained on data from three temperatures and tested on an unseen intermediate temperature (either 272°C or 450°C), ensuring no data leakage from the test set into the training, PCA decomposition, or grid construction processes.

Key Results
The JC-PCA-MLP model demonstrated superior performance compared to the standalone JC model, a pure MLP, and a PCA-MLP model without a physical baseline:

At 272°C (Unseen): The model achieved an average $R^2$ of 0.9306 and an Average Absolute Relative Error (AARE) of 1.97%. The minimum $R^2$ was 0.8779.
At 450°C (Unseen): The model achieved an average $R^2$ of 0.9538 and an AARE of 1.63%. The minimum $R^2$ was 0.8909.
Comparative Performance:
- The standalone JC model performed poorly on unseen temperatures (Average $R^2$ < 0.45), exhibiting significant systematic deviations.
- The pure MLP model showed instability, with performance degrading significantly at 450°C (Average $R^2$ = 0.2386).
- The PCA-MLP model (without JC baseline) improved upon the MLP but still lacked the stability and accuracy of the hybrid approach, particularly in maintaining high minimum $R^2$ values.

Significance and Claims
The paper claims that the JC-PCA-MLP framework offers an efficient physics-informed data-driven strategy for dynamic constitutive modeling. Its significance lies in the synergistic combination of three elements:

Physical Guidance: Using the JC model as a baseline ensures the prediction adheres to fundamental physical trends, reducing the learning burden on the neural network.
Residual Learning: By focusing the MLP on correcting the JC model's systematic errors rather than learning the entire curve, the model achieves higher accuracy with limited data.
Dimensionality Reduction: The use of PCA imposes global constraints on the curve shape, preventing the local oscillations and overfitting common in point-wise neural network predictions.

The authors conclude that this method effectively enhances prediction accuracy and stability for AM 316L stainless steel under high-temperature and high-strain-rate conditions, providing a viable approach for constitutive modeling where experimental data is scarce. They suggest this strategy could serve as a reference for flow stress prediction in other additively manufactured metallic materials.

More like this