Abstract
Aiming at the problem of high cost of mechanics performance testing and uncontrollable experimental accuracy, a new "mapping model" based on Support Vector Machine (SVM) is proposed. The SVM regression nonlinear prediction model was established using a small sample of experimental data, taking temperature as the input and mechanical properties (elastic modulus, ultimate strength, yield strength) as outputs. Stress-strain properties were analyzed using MATLAB, and constitutive relations suitable for engineering use were obtained. Results show that SVM-based models capture complex nonlinear relationships with lower relative error and higher goodness of fit compared to traditional neural networks.
Introduction
In structural engineering, Q235 steel is a foundational material. However, its mechanical properties—specifically its elastic modulus, yield strength, and ultimate strength—fluctuate significantly with temperature. Traditionally, determining these properties requires extensive physical tensile testing, which is costly, time-consuming, and difficult to control for accuracy across wide temperature ranges. This research addresses the need for a mathematical model that can accurately predict material behavior from limited experimental data.
Methods and Materials
The study utilizes a machine learning approach to establish a constitutive relation for Q235 steel.
Experimental Baseline
Tensile tests were conducted on 23 specimens across 11 temperature points ranging from -60°C to 600°C. These tests provided the "ground truth" data necessary to train and validate the model.
SVM Regression Modeling
A Support Vector Machine (SVM) was employed for nonlinear regression. Unlike standard neural networks that require large datasets, SVMs excel with small samples by identifying support vectors that define an optimal regression hyperplane. An exponential kernel function was used to map input data into a higher-dimensional space, facilitating the modeling of complex nonlinearities. All variables were normalized to a [0, 1] range to ensure model stability.
Key Insights
The research highlighted several critical observations:
- Small Sample Efficiency: SVMs proved superior to Back-Propagation (BP) neural networks for engineering problems with limited data, avoiding the overfitting common in small-scale neural network training.
- Blue Brittleness: The model successfully accounted for the "blue brittle" phenomenon observed at 300°C, where the steel's yield platform becomes less distinct.
- Constitutive Equations: By integrating experimental and forecasted data, high-order polynomial equations were derived, allowing engineers to estimate material properties directly from temperature inputs.
Key Results
The SVM model demonstrated exceptional precision. For instance, the relative error for the elastic modulus at -40°C was a mere -0.003%.
As shown in the error analysis, the SVM's relative error remained consistently lower and more stable than the BP neural network across the entire temperature spectrum for elastic modulus, yield strength, and ultimate strength. The study confirmed that while strength and elasticity generally decrease as temperature rises, the SVM model captures the specific rates of decline with high fidelity.
Conclusion
The proposed SVM-based mapping model provides a high-precision, low-cost alternative to exhaustive physical testing. While currently calibrated for Q235 steel, the methodology offers a universal framework for modeling complex material behaviors in extreme environments, potentially streamlining safety checks in structural design software.
Summary updated