Enhancing the high-temperature constitutive modeling of In706 superalloy using advanced artificial neural networks and attentive staged optimization algorithm

Abstract

High-temperature constitutive models derived from the Arrhenius equation play a pivotal role in the hot-forging processes of nickel-based superalloys such as In706. However, existing models have limitations owing to insufficient data points, leading to the potential loss of vital information and the need for elaborate modeling methodologies. This study proposes an innovative optimization approach that leverages the attentive staged optimization algorithm (ASOA) and advanced artificial neural networks to enhance the accuracy and efficiency of In706 superalloy modeling. ASOA is a batch-based, multi-stage optimization algorithm inspired by the attention mechanism concept in deep learning. It allocates time and computational resources to the most deserving data for optimization and is well-suited for handling large volumes of data. In this paper, ASOA was employed to perform two rounds of fitting on the thermal simulation experimental data, resulting in the determination of the critical strain ϵc and identifying suitable Arrhenius equation material parameters for each set of deformation conditions. Subsequently, a full parameter responsive constitutive model was established for In706, capable of handling a comprehensive set of deformation conditions and providing a more detailed representation of the high-temperature behavior of the In706 alloy. In developing the FPR model, we designed a fully connected neural network and employed a cosine annealing dual-error dynamic adjust function to address error amplification during training. The cosine annealing learning rate scheduling strategy with amplitude attenuation was also implemented to dynamically adjust the learning rate. As a result, the model developed using the proposed approach outperforms traditional models and shows remarkable potential for the precise and efficient optimization of In706 alloy. The high performance of the final model showcases the prospect of our approach in future development and optimization of superalloys.