基于人工神经网络的高精度TA16钛合金本构模型

High-Accuracy Constitutive Model of TA16 Titanium Alloy Based on Artificial Neural Networks

摘要:

使用Gleeble-3800对锻态TA16钛合金在温度730~1030 ℃、应变速率0.1~10 s^-1的条件下进行热模拟压缩实验，取得TA16合金在该变形条件范围内的真应力-真应变曲线。分别使用Arrhenius、Johnson-Cook本构模型和人工神经网络（ANN）3种方式建立了TA16钛合金的本构模型，并对模型误差进行了分析。结果表明：TA16钛合金在中、低应变速率下屈服后加工硬化与软化达到动态平衡状态，在高应变速率下呈现先软化后再进入平衡状态，合金加工性能良好；Arrhenius、Johnson-Cook和ANN建立的TA16钛合金本构模型平均绝对百分比误差（MAPE）分别为11.49%、23.7%和1.69%，ANN模型较传统本构模型精度高1个数量级；Arrhenius本构模型在中、高应变速率和中、低应变范围内精度较好，在工程中具有实用性；Johnson-Cook本构模型体现了高应变硬化的趋势，难以描述TA16钛合金屈服后动态平衡状态，模型精度较差，不宜在工程中使用；ANN本构模型在全部实验参数范围内具有极高的预测精度，同时在实验参数以外预测数据同样具有良好的精度，能够为工程实践提供高精度的本构模型。

Abstract:

The thermal simulation compression experiments were conducted on forged TA16 titanium alloy using the Gleeble-3800 system at the temperature of 730–1030 ℃ and strain rates from 0.1 s^–1 to 10 s^–1. The true stress-true strain curves of TA16 alloy under these deformation conditions were obtained. Constitutive models for the TA16 alloy were established using three different methods: the Arrhenius model, the Johnson-Cook model, and artificial neural networks (ANN). The model errors were analyzed. The results indicate that the TA16 alloy reaches a dynamic balance between work hardening and softening after yielding at medium and low strain rates. At high strain rates, it initially softens and then enters a balance state, demonstrating good workability. The mean absolute percentage error (MAPE) of the constitutive models for the TA16 alloy using the Arrhenius model, the Johnson-Cook model, and ANN is 11.49%, 23.7%, and 1.69%, respectively. The ANN model shows an order of magnitude higher accuracy compared to the traditional constitutive models. The Arrhenius model exhibits better accuracy at medium and high strain rates and in the medium and low strain range, making it practical for engineering applications. The Johnson-Cook model reflects the trend of high-strain hardening and struggles to describe the dynamic equilibrium state after yielding for the TA16 alloy, resulting in poor model accuracy and making it unsuitable for engineering applications. The ANN model demonstrates extremely high predictive accuracy across the entire range of experimental parameters, and it also maintains good accuracy for data predictions beyond the experimental parameter ranges, providing a highly accurate constitutive model for engineering applications.