A New Perspective of Post-Weld Baking Effect on Al-Steel Resistance Spot Weld Properties through Machine Learning and Finite Element Modeling
Round 1
Reviewer 1 Report
This paper shows two interesting works to study the post-baking effect of the RSW of Al/steel, one is a ML framework, the other one is a FEM modelling. Several interesting founding are provided. However, this paper still needs improvements.
1. From my point of view, the main problem of this paper is that the two parts of the work have no real connection with each other. If we only check the experimental results, the ML model is even not needed to get the main concluion of the first part: "post-weld baking reduces the joint performance, and the extent of degradation is inversely proportional to the thickness of the steel sheet within the stack-up". More analysis from the ML should be provided.
2. Is a stress free condition used as the intial condition of the FEM model? Usually, the weld directly from a RSW process is not stress-free.
Author Response
We sincerely appreciate the valuable comments and suggestions from the two reviewers. We have carefully revised our papers and prepared the following replies. The original review comments are given in blue.
This paper shows two interesting works to study the post-baking effect of the RSW of Al/steel, one is a ML framework, the other one is a FEM modelling. Several interesting founding are provided. However, this paper still needs improvements.
1. From my point of view, the main problem of this paper is that the two parts of the work have no real connection with each other. If we only check the experimental results, the ML model is even not needed to get the main conclusion of the first part: "post-weld baking reduces the joint performance, and the extent of degradation is inversely proportional to the thickness of the steel sheet within the stack-up". More analysis from the ML should be provided.
We thank the reviewer’s comments. The present work utilized the machine learning and finite element modeling to study the post-weld baking effect on resistance spot welds of Al and steel alloys. We first agree with the reviewer that any properly constructed and trained machine learning tool can only reveal the correlation or effects that exist in the experiment. ML can’t reveal what is not in the experiment. Second, we agree with the reviewer that, for an experienced researcher who have intimate knowledge of the specific welding experiment such as single factor experiment involving the bake painting effect, the effect of reduced joint strength is observable. However, such an effect may not be obvious, or buried with other factors, when dealing with many different material and welding combinations such as the huge amount of dataset in this study. This is where machine learning, when a ML model is properly structured and trained, would be powerful in identifying such effect among complex interactions among many variables. The main objectives of employing the two approaches are to:
- Our ML model was able to establish the relationship between many weld variables and joint performance, and the ML model was able to reveal the post-weld baking effect and many other effect from the huge dataset used in the study. The post-weld baking effect was highlighted in present work, as it appeared to be contradictory to the general understanding of the baking effect discussed in other’s studies, such as those cited in the paper. Through iterative training, the ML model was able to construct the high dimensional correlations between the weld variables and joint performance (as indicated by the high correlation coefficient between the predicted and observed joint performance properties shown in Fig.2). Meanwhile, the effect of post-weld baking on joint performance was quantified by the ML model (see Fig. 3, the ML predicted baking induced joint performance exhibited quantitative agreement with the experimental measurements for welds with multiple material thickness combinations). The fully trained ML model is very beneficial to provide quantitative prediction for weld performance of new ‘untested’ materials and thickness combinations.
- While the ML model quantified the baking induced effect on joint performance, we are still unclear about the underlying physical mechanism. The finite element modeling provided us the fundamental understanding of the underneath root cause, that is, the formation of high thermal stresses at the faying interface caused by the mismatch of thermal expansion strain between steel and Al alloy. The physical understanding is helpful for guiding welding process development to mitigate the negative effect of post-weld baking on dissimilar Al-steel welds.
The present work suggests that ML would be more powerful when it is combined with other research tools such FEM analysis. The combinations are essential, so that not only we can quantify the roles of certain process variables on the joint performance, but also develop the physically understanding the post-weld baking effects.
2. Is a stress free condition used as the initial condition of the FEM model? Usually, the weld directly from a RSW process is not stress-free.
Yes, we assumed the stress-free condition as the initial condition of FEM model. We agree that the welds would have residual stresses due to the thermal-mechanical process through resistance spot welding, and the magnitude and distribution of residual stresses could vary in welds made of different materials and thicknesses. Since we focused on the effects induced by post-weld baking process, the stress free condition was assumed as the initial state of the welds, so that the post-weld baking effect on welds of different thickness combinations can be directly compared.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The topic taken up in the article seems quite interesting and worth publishing, but a few issues seem to be important to clarify. The goal to be achieved by the author, as they declare in the abstract, is steel dissimilar resistance determinantion using machine learning (ML) and finite element modeling.The optimization is declared to be carried out via deep neural network. In my humble opinion, some aspects are not very clear. More in details:
Lines 32-35 - Please explain how joining aluminum to steel affects greenhouse gas emission and vehicle fuel economy.
Line 52 - Please expand the abbreviation SEM.
Line 53 - state what intermetallic compounds are formed.
Figure 1 – Explain “BM”
Figure 1 - Please explain in more detail the parameters found in the neural network in the input and output layers. Please write:
- definition / meaning of each parameter,
- how it was measured, with what measuring device.
Line 135 - What is Minimum-Maximum normalization?
Line 164 – What exactly did the authors mean by these tasks - “The experimental data was standardized and transformed into readable formats”
The description of the structure of the neural network is not sufficient. Some key information is missing:
- how many layers of hidden networks are there,
- what are the values of the weights of interneural connections
- what neuron activation functions were used in the network
- what neural network learning algorithm was used
Similar issues have been described in other publications. Authors can find information, for example, in the following papers, Maybe worth mentioning in the description and references:
https://doi.org/10.3390/app112110414
Lines 226-227 – “Both the experimental measurement results and the ML predictions are presented in the figure” – in which figure?
Author Response
We sincerely appreciate the valuable comments and suggestions from the two reviewers. We have carefully revised our papers and prepared the following replies. The original review comments are given in blue.
The topic taken up in the article seems quite interesting and worth publishing, but a few issues seem to be important to clarify. The goal to be achieved by the author, as they declare in the abstract, is steel dissimilar resistance determination using machine learning (ML) and finite element modeling. The optimization is declared to be carried out via deep neural network. In my humble opinion, some aspects are not very clear. More in details:
Lines 32-35 - Please explain how joining aluminum to steel affects greenhouse gas emission and vehicle fuel economy.
Joining aluminum to steel alloys is utilized by automotive industries for lightweighting of vehicle body structure which is one of the main factors affecting the vehicle fuel economy and greenhouse gas emissions.
Line 52 - Please expand the abbreviation SEM.
Here SEM stands for scanning electron microscope.
Line 53 - state what intermetallic compounds are formed.
The intermetallic compounds in Al-steel resistance spot welds consisted of two phases, one is a blocky Fe2Al5 phase adjacent to the steel substrate and the other is a thin layer of FeAl3 adjacent to the Al substrate.
Figure 1 – Explain “BM”
BM stands for the base material of steel and Al alloys.
Figure 1 - Please explain in more detail the parameters found in the neural network in the input and output layers. Please write:
- definition / meaning of each parameter
- how it was measured, with what measuring device.
In the Input layer, there are 4 major types of input variables, including
Weld attributes:
- Button sizes: defined as the dimensions measuring the length across the button retained on the post fractured specimens; they were measured by metallographic tests
- Material indentation: it was formed due to pressure and heat during RSW and measured by metallographic tests
- Expulsion: it was formed due to excessive heat during RSW and measured by metallographic tests
- Nugget size: it is the diameter of weld nugget that was solidified from the melt pool of metals
- IMC thickness: thickness of IMCs formed between Al and steel
- Hardness: hardness measuring the material strength across the weld
Base materials:
- The thickness of Al and steel sheets
- Type of Al and steel alloys
- Type of coating on steel alloy
Coupon geometry:
Geometry dimensions of the tested weld coupon, as presented in Fig. 4(a).
Other conditions:
- They included the weld fabrication conditions, such as the adhesive, post-weld baking, aging conditions, and the fit-up conditions (for example misalignment and gap between Al and steel sheets) of stacking sheets as well as palm print materials (material sheet is not flat).
In the Output layer, they are peak load, extension at break, and total energy measured under coach peel tests (test coupon shown in Fig. 4(a)). The three properties were selected to represent the weld’s strength, elongation, and resistance to fracture, respectively.
Line 135 - What is Minimum-Maximum normalization?
The Minimum-Maximum normalization is to rescale the data streams in the range of [0, 1] using the following equation (please see the attached pdf version), where represents the ith data stream, and denote the minimum and maximum of the data stream, and is the corresponding normalized data. The normalization helps to diminish the scale difference between different data streams and improve training stability by encouraging more balanced weights of different data streams.
Line 164 – What exactly did the authors mean by these tasks - “The experimental data was standardized and transformed into readable formats”
The experimental data was collected through years of research and consisted of inconsistent data formats and records (for example, different symbols were used for recording the same experimental observations). To address such situation, we made efforts to prepare, standardize, and transform the experimental data into the formats that can be directly used for training and testing the ML model.
The description of the structure of the neural network is not sufficient. Some key information is missing:
- how many layers of hidden networks are there,
- what are the values of the weights of interneural connections
- what neuron activation functions were used in the network
- what neural network learning algorithm was used
We utilized three hidden layers in the neural networks. Following each layer, a dropout layer was applied to regularize the neural network in order to prevent overfitting and improve the model’s generalization capability. The weight constraint was imposed on the hidden layers to force the magnitude of the weights and avoid large weights on neurons that may cause the risk of overfitting and difficulty of convergence. The rectified linear activation function was used in neural network and Adam optimization algorithm was applied to update network weights. The design of architecture, activation function, and learning algorithm were selected based on the model performance by grid searching a variety of trials. More details of model design and description of input parameters (mentioned in comment #5) will be discussed in a coming paper that is focused on methodology development. The present paper is more on utilizing the ML model to illustrate the post-weld baking effect on joint performance.
Similar issues have been described in other publications. Authors can find information, for example, in the following papers, Maybe worth mentioning in the description and references:
https://doi.org/10.3390/app112110414
We thank for the reviewer’s suggestion and have cited this reference in the revised version.
Lines 226-227 – “Both the experimental measurement results and the ML predictions are presented in the figure” – in which figure?
It is Figure 3.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
This manuscript can be accepted in the current form.
Reviewer 2 Report
thank you for the corrections made, the article has certainly benefited from the corrections made
