PSI - Issue 51

Kenichi Ishihara et al. / Procedia Structural Integrity 51 (2023) 62–68 K. Ishihara and T. Meshii / Structural Integrity Procedia 00 (2022) 000–000

63

2

Nomenclature v

Weld speed

Heat source power

Q

a , b , c , F

Parameters of Goldak’s double-ellipsoidal heat source model

W , D

Width and Depth of molten shape

Density

ρ λ

Thermal conductivity

Specific heat

c p

Thermal diffusivity

κ

Heat flux

q

x , y , z T , T 0

Cartesian coordinates

Temperature, and initial temperature

Time from welding start

t

Abbreviations ANN

Artificial neural network

DL

Deep learning

FEA TDA

Finite element analysis

Transient temperature distribution analysis

to accurately reproduce the transient heat input in the welding procedure, which usually differs from that planned in the welding procedure. Thus, traditionally, inverse analysis on the transient heat input is conducted by focusing on the molten shape of the actual components; i.e., try and error work performing transient temperature distribution analysis (TDA) assuming heat input until the molten shape is reproduced. There were many heat input models used for welding TDA, but recently most of the researchers are using Goldak’s double-ellipsoidal heat source model (Goldak, 1984). In this case, heat source parameters ( a , b , c ) from experience are assumed, and if the TDA results and the molten shape are different, TDA is repeated by changing the parameters ( a i , b i , c i ), here the subscript i indicates the number of iterations, until the obtained region above the melting point matches the target molten shape. When a valid TDA results are obtained, the residual stress in the welds can be calculated accurately by conducting a thermal stress analysis using obtained the TDA results. As can be imagined, it took an excessive amount of time to adjust the heat source model parameters ( a , b , c ), which was a bottleneck in the work.

Fig. 1.Engineering framework to inverse analysis method of molten shape by deep learning. (Image figure of welded metal cross-section (Chujutalli, 2016)).

Recently, machine learning methods have been become effective to change inverse analysis in structural integrity issues to straight-forward analysis (Ishihara, 2021). Thus, in this work, a model using the deep learning (DL) has been developed that can directly determine the parameters ( a , b , c ) of the heat source model from the welding records and the molten shape, which can omit the inverse analysis of the heat source parameters, which has been a bottleneck in