Exploration of physics-related latent vectors in hot working of Inconel 718 superalloy using autoencoder

The thermo-mechanical processing (TMP) behavior of nickel-based superalloys during hot working exhibits highly non-linear and complex characteristics, necessitating an understanding of deformation mechanisms for industrial applications. In this study, we developed an autoencoder-prediction network (AE-PN) to model the TMP behavior of the Inconel 718 superalloy in an unsupervised way within the temperature range of 900-1200 °C and strain rates of 0.001-10 s -1 . The AE-PN effectively reduces the dimensionality of stress-strain curves in a non-linear manner while preserving critical flow features, enabling accurate reconstruction of stress-strain curves. The AE-PN established strong correlations between processing parameters and the latent space, achieving high reconstruction accuracy for continuous flow curves (testing RMSE = 4.19). Latent features derived from stress-strain curves were linked to key characteristics of hot deformation behavior, including peak stress, strain hardening, and flow softening. Notably, one major latent component was strongly correlated with peak stress ( R 2 = 0.9971), highlighting its role as a physics-related variable connected to TMP behavior. For additional datasets within the processing window, the AE outperformed a conventional artificial neural network (ANN) model, achieving an RMSE of 7.98 compared to the ANN's RMSE of 60. Furthermore, hot workability characteristics, such as flow softening and dynamic recrystallization (DRX), were systematically analyzed using a processing map and microstructural analysis. This AE-PN framework provides interpretable insights beyond traditional ‘black-box’ systems, offering accurate stress-strain curve predictions and deeper interpretation of processing parameters through latent space exploration. • Development of a novel deep learning framework, namely “autoencoder-prediction network (AE-PN)” to predict and model thermo-mechanical processing (TMP) behavior for Inconel 718 superalloy with high accuracy for testing subset (RMSE = 4.19). • Unsupervised learning-based AE’s capability of reconstruction is superior over traditional artificial neural network (ANN) models in additional datasets within TMP window (RMSE: 7.98 vs 60). • The latent components derived from stress-strain curves without labeled information were associated with key characteristics of TMP response (e.g. peak stress, strain hardening, flow softening). Notably, the specific latent component showed a robust correlation with peak stress ( R 2 = 0.9971), including supplementary datasets. • The hot workability of Inconel 718 alloy was systematically analyzed through the development of processing map and microstructural observation, including hot deformation mechanisms such as flow softening, dynamic recrystallization (DRX).