Uncertainty quantification in multivariable regression for material property prediction with Bayesian neural networks.

With the increased use of data-driven approaches and machine learning-based methods in material science, the importance of reliable uncertainty quantification (UQ) of the predicted variables for informed decision-making cannot be overstated. UQ in material property prediction poses unique challenges, including multi-scale and multi-physics nature of materials, intricate interactions between numerous factors, limited availability of large curated datasets, etc. In this work, we introduce a physics-informed Bayesian Neural Networks (BNNs) approach for UQ, which integrates knowledge from governing laws in materials to guide the models toward physically consistent predictions. To evaluate the approach, we present case studies for predicting the creep rupture life of steel alloys. Experimental validation with three datasets of creep tests demonstrates that this method produces point predictions and uncertainty estimations that are competitive or exceed the performance of conventional UQ methods such as Gaussian Process Regression. Additionally, we evaluate the suitability of employing UQ in an active learning scenario and report competitive performance. The most promising framework for creep life prediction is BNNs based on Markov Chain Monte Carlo approximation of the posterior distribution of network parameters, as it provided more reliable results in comparison to BNNs based on variational inference approximation or related NNs with probabilistic outputs.

The Effect of Welding Defects on the Long-Term Performance of HDPE Pipes.

High-density polyethylene (HDPE) pipes are the preferred pipes of water systems in nuclear power plants because they are durable, corrosion-free, easy to install, and not subject to fouling. However, their long-term performance can be affected by welding defects. In this paper, the effect of welding defects on the long-term performance of HDPE pipe butt fusion joints was studied using a creep test. A welding defect with a hole or inclusion in the joint was simulated by artificially inserting a copper ball during butt fusion welding. The test results showed that the creep life of the joint decreased with increased defect size. An expression describing the creep life and the defect ratio was obtained according to the test results. In addition, the test results showed that the creep life of the joint without a welding bead was about 50% of that in a joint with a welding bead.

Influence of Post-Bond Heat Treatment on Microstructure and Creep Behavior of the Brazed Single-Crystal Nickel Superalloy.

Post-bond heat treatment (PBHT) is an effective way to improve the bonding quality of a brazed joint. Herein, brazing of a nickel-based single crystal superalloy is carried out using a Ni-Cr-Co-B-Si-Al-Ti-W-Mo filler alloy, and the microstructure and creep property of the brazed joint are systematically investigated using scanning electron microscopy (SEM), Thermo-Calc software, an electron probe micro-analyzer (EPMA), X-ray diffractometer, confocal scanning laser microscope (CSLM), and transmission electron microscopy (TEM). The results reveal that the as-prepared joint only consists of an isothermally solidified zone (ISZ) and an athermally solidified zone (ASZ), where the cubic γ' phase is observed in the ISZ, and skeleton-like M3B2, γ + γ' eutectic and reticular G phases are observed in the ASZ. Furthermore, the γ + γ' eutectic and G phases disappear and the M3B2 alters from a skeleton-like to block-like shape in the ASZ after PBHT. Meanwhile, some lath-like M3B2 phases are precipitated at the edge of the ISZ and several M3B2 phases are precipitated in the base metal, forming a new zone in the brazed joint, namely at the diffusion affected zone (DAZ). Owing to the removal of low melting point eutectics from the as-prepared joint, the creep life also increases from 188 h to 243 h after PBHT. The current work provides a method for the optimization of brazed joints based on the Ni-based single crystal superalloy.

Creep Life Prediction of 10CrMo9-10 Steel by Larson-Miller Model.

Creep is defined as the permanent deformation of materials under the effect of sustained stress and elevated temperature for long periods of time, which can essentially lead to fracture. Due to very time-consuming and expensive testing requirements, existing experimental creep data are often analyzed using derived engineering parameters and models to predict and find the correlations between creep life (time to rupture), temperature and stress. The objective of this study was to analyze and compare different numerical algorithms by using the Larson-Miller parameter (LMP) extrapolation model. Calculations were performed using the classical LMP equation where different values of parameter C were selected, as well as using a modified LMP equation in which parameter C was stress dependent C(σ). The impact of two different approaches of extrapolation and correlation functions (linear and polynomial) applied to fit the LMP model was also investigated. A detailed analysis was performed to choose the best extrapolation fit function and error tolerance. The numerical algorithm implemented was validated through creep rupture testing performed on 10CrMo9-10 steel at 600 °C (873 K) and 80 MPa. Creep model behavior analysis proved that different values of C can significantly change the estimated time to rupture. An excellent response of the LMP model was obtained by considering polynomial dependency when parameter C was assumed to be 18, especially for the temperature range from 773 to 873 K. Promising results were also achieved when parameter C was taken as stress-dependent, but only for linear fitting, which requires further analysis. However, at validation stage it turned out that only the linear extrapolation function and C taken as a constant value provided adequate time-to-rupture prediction. In the case of C = 18, estimated time was slightly overestimated (~8%) and for C = 20 it was underestimated by 27%. In all other cases error largely exceeded 50%.