Detection of Hydrogen Bubbles Produced by Corrosion Inhibition of Metal Weldment Using Machine Learning

Abstract

This study analyzes the corrosion inhibition efficiency of sodium molybdate (SM) solution on weldment specimens in 1 M HCl, based on H₂ bubbles detection. The detection of the bubbles produced by the corrosion process is achieved by a YOLOv4 deep learning algorithm. The results indicate that the corrosion rate is higher on the weld metal zone than on the base metal zone in the same solution, which might be attributed to the coarser grain of the weld metal zone and the stability of the produced oxide layer. The addition of sodium molybdate was discovered to boost the stability of the oxide layer, hence enhancing the specimens’ corrosion resistance. The overall inhibitory efficiency of the sodium molybdate solution was 59% for the weld metal (WM) (0.4 g/L), 52% for the heat-affected zone (HAZ) (0.2 g/L), and 37% for the base metal (BM) (0.2 g/L). The object detection algorithm models showed 97% for the mAP and 0.98 for recall. The minimum average bubble detected for the WM was 0.353 /mm² at an SM concentration of 0.4 g/L, while the HAZ was 0.612 /mm² at 0.2 g/L, and the BM was 1.055 /mm² at 0.2 g/L. The results of the bubbles detection appeared to be consistent with the corrosion experiment outcomes obtained by the potentiodynamic polarization and hydrogen volume measurement tests. This experiment validates the distinctiveness of the different weld zones in terms of the inhibitor concentration required for successful corrosion prevention, as well as the potential of analyzing corrosion using machine learning algorithms for object detection.

1. Introduction

The application of corrosion inhibitors is among the most useful techniques to preserve the internal surface of steel structures, for example, tanks and pipelines, for facing corrosion, mainly in acidic environments. An acidic environment can cause various problems with metals, for example, stress corrosion cracking (SCC) in chloride-rich environments [1]. Increasing a tiny number of inhibitors without substantially varying the concentration of any corrosive agent can decrease and avoid the reaction of a metal to its environment. The benefits of inhibitors are that they reduce the level of corrosion between the corrosive agents and the surface of the metal by making a passive film on the metal surface and do not intervene in the mechanical properties of the material compared to other corrosion prevention techniques [2,3]. Organic materials have great prospects for maximizing their role in the field of engineering [4], including for modifying corrosion processes. Corrosion inhibitors can be produced from organic and non-organic compounds. The advantages of organic compounds are they are non-toxic, affordable, renewable, and easy to produce [5,6] but still need a lot of research to fully understand their performance and mechanism. On another side, non-organic inhibitors have already been studied and tested [7,8,9].

The weldment metal exhibits microstructural variations along the melted zone/weld metal (WM), heat-affected zone (HAZ), and base metal (BM) because of the welding parameters [10] and process itself [11], and it is also able to cause galvanic corrosion in steel [12]. Welding parameters significantly influence the welded joints’ quality, structural integrity, and mechanical properties. Variations in parameters, such as the current type and magnitude, voltage, travel speed, electrode specifications, and shielding gas composition, directly impact crucial aspects, like the penetration depth, heat input, bead geometry, and the distribution of residual stress [10,13,14,15]. Achieving optimal parameter settings is essential to attain the desired weld properties and prevent defects. Studies on shielding gas show the effects of the shielding gas content on the microstructure and toughness, underscoring the importance of precise parameter control in welding processes [16]. Differences in the microstructure, including the grain size and shape, impact the active–passive response of a corroded surface in a welded structure. For instance, a finer grain may result in a reduced localized corrosion rate but an increased rate of uniform corrosion [17]. The corrosion behavior of the HAZ region of welded X70 steel pipes in naturally aerated solutions containing 0.5 M bicarbonate and 0.01 M chloride at an atmospheric temperature of 25 °C have been investigated by Khalaj et al. [18]. They found that the bainitic microstructure generated in the area that experiences fast cooling rates and high temperatures during welding such as the coarse grain HAZ (CGHAZ) and partially melted zone (PMZ) was easy to corrode as showed by their higher current density, whereas the weld metal region had the lowest corrosion current density.

Several investigations within welding research have underscored the presence of solid phase transitions that notably affect stress modifications during welding [19,20,21,22]. This transition commonly aligns with changes in mechanical characteristics, volumetric expansion, and Transformation-Induced Plasticity (TRIP). These modifications deeply impact the formation of residual stresses, prompting a recent focus on incorporating the effects of microstructure transformation into both experimental and simulated thermal–mechanical models for welding [23,24,25,26].

The mechanical properties of each weldment zone are significantly influenced by hydrogen, with the hydrogen diffusion behavior variance and stress field distribution in each phase being crucial factors impacting hydrogen embrittlement behavior [27]. Hydrogen atoms exhibit a tendency to diffuse and accumulate around imperfections in the crystal lattice, leading to elevated hydrogen pressure [28,29]. This process reduces the bond energy [30,31] and aids in the accumulation of dislocations or localized plastic deformation [32,33].

The main purpose of the inhibitor is to control hydrogen evolution and corrosion. This is performed when the reaction with the metal is taking place and the water decomposition process occurs [34,35]. The inhibitor must assist in the recombination of hydrogen gas and inhibit the penetration of hydrogen into the metal lattice to avoid embrittlement [36]. The detection of hydrogen gas produced during the corrosion inhibition process could be used as an indication of inhibitor performance [37].

Standard electrochemical test methods such as potentiodynamic polarization (PP) can be applied to determine inhibitor effectiveness in corrosion inhibition tests [38]. To gain more insight into inhibitor effectiveness, a combination approach can be performed using the detection of hydrogen gas (H₂ bubbles) with these electrochemical methods and the bubble detection. This approach is able to provide more information on the effectiveness of an inhibitor.

Image extraction of the corrosion process could be performed through a polarization test, based on video acquisition. Each image provides insight into the metal surface material transformation process and the H₂ bubbles’ spread [34,38]. Detecting the H₂ bubbles at a certain corrosion rate condition is not an easy job, for instance, in scenarios characterized by low current density [34,39,40]. To detect bubbles from a corrosion process in a metal weldment image, a machine learning algorithm (GLCM/SVM method) was implemented [6]. In the deep learning concept, the computer will attempt to map a feature from a set of images through a training process to acquire certain patterns, for example, object sizes, color channels, and so on, to support object detection schemes [41,42]. Presently, a promising technology for object detection is YOLO (You Only Look Once), which has several parameters for justification, including the resulting model size, detection speed, and detection accuracy [43]. This algorithm is one of the popular object detections that has advantages in real-time performance and must be implemented with a training process on image data with the expected object class [44,45]. The computation of these algorithms can be run by utilizing the computer’s main processing tools or with the help of a graphics card to generate models for object detection. After the bubble is detected by the algorithm, the bubble can be quantified based on the weldment zones (weld metal, heat-affected zone, and base metal). Machine learning algorithms have been involved in the study of corrosion and inhibitors, such as quantum chemical variables [46], the molecular dynamics of the inhibitors [47], and measuring the immersion time and corrosion rate [48].

The purpose of this work is to detect bubbles produced from the corrosion inhibition process of a sodium molybdate on carbon steel based on H₂ evolution using a deep learning approach. The integration of conventional electrochemical testing and advanced deep learning algorithms in this research is a novel approach that has the potential to improve the field. This novel combination not only enhances the accuracy of the results but also opens up new avenues for further research and discovery. By calculating the corrosion rate based on PP tests on carbon steel specimens in 1 M HCl solution with up to 0.4 g/L sodium molybdate and detecting bubbles with a deep learning method on hydrogen gas bubble images captured during the electrochemical test, this work reveals some fundamental aspects of the inhibition mechanism and provides an example of the usefulness of deep learning in corrosion inhibitor research.

2. Materials and Methods

The process flowchart of the research framework is provided in Figure 1.

2.1. Weldment Preparation

The weldment specimen was prepared using a Shielded Metal Arc Welding (SMAW) technique and E6013-type electrodes on AISI 1040 carbon steel plates (0.38 wt% C and 0.6 wt% Mn). The AWS A5.1 standard was employed [49], where the plate thickness is 8 mm. The SMAW technique was carried out due to its versatility and effectiveness in welding thicker steel plates, which is crucial for this project’s structural requirements. Its ability to penetrate and create strong, reliable welds on heavy materials, along with its adaptability to various environmental conditions, made it the optimal choice for ensuring the robustness and durability of the welded joints in this specific construction. The welding parameters are listed in

Table 1. SMAW process parameters.

Parameters	Values
Current Type	DC
Polarity	DCEP (Direct Current Electrode Positive)
Current (A)	90
Electrode Type	E6013
Arc Voltage (V)	24
Travel Speed (mm/min)	±150
Electrode Angle (degrees)	±45
Arc Length (mm)	3–8
Type of Butt Welds	Double-V butt joints with single pass each side
Groove Angle (degrees)	60
Butt Gap (mm)	1

.

2.2. Electrochemical Corrosion Testing

The corrosion testing methods for the specimens included hydrogen collection measurement, electrochemical testing, and hydrogen bubble detection. The specimens for the corrosion testing were divided into two types. The first type was cut from the weldment samples at the three zones of interest and ground with abrasive papers up to 2000 grit, and the second type used resin and was mounted in a container for electrochemical testing and recorded for video. The first type was applied to hydrogen collection and electrochemical testing and the second type targeted hydrogen bubble quantification.

Sodium molybdate (Na₂MoO₄) in powder form (22.3 wt% Na, 46.6 wt % Mo, and O 31.1 wt %) was dissolved into a 1 M HCL electrolyte solution. The solution will act as a corrosion test solution with varying concentrations of 0 g/L, 0.1 g/L, 0.2 g/L, 0.3 g/L, and 0.4 g/L.

A range of corrosion testing solutions were prepared with various percentages of sodium molybdate, Na₂MoO₄ inhibitors, i.e., 0.1, 0.2, 0.3, and 0.4 g/L, into 1 M HCl. The electrochemical test methods were performed by potentiodynamic polarization with an Autolab potentiostat (PGSTAT 128N, Herisau, Switzerland). A three-electrode system was used for conducting the PP test, with the weldment specimen as the working electrode as shown in Figure 2a. The reference electrode was Ag/AgCl (3 M KCl), and the counter electrode was platinum. The PP test was measured at +/− 1 V from the corrosion potential (E_corr) at a scan rate of 1 mV/s. Then, the Tafel plot extrapolation was analyzed by using the software tool associated with the potentiostat NOVA 1.11.

The inhibition efficiency IE (%) of the inhibitor was calculated by using the equation below:

$$IE\left(\%\right)=\frac{i_{corr}-i_{corr\left(inh\right)}}{i_{corr}}$$

where i_corr(inh) and i_corr are the corrosion current densities with and without the inhibitor, obtained by the extrapolation of the Tafel lines to the corrosion potential from the PP test.

Figure 2b shows the minicell was linked to the three-electrode cell within the solutions contained in the specimen container [6]. After closure, it was directed toward a 50 mL syringe designed for dissolved gas analysis (DGA) (Figure 2b), as well as to the solutions. This configuration was employed to quantify the volume of H₂ gas generated during the PP test.

Three techniques—potentiodynamic polarization, hydrogen volume measurement, and deep learning—were employed using five specimens. Each test specimen underwent evaluation in distinct types of HCl solutions: HCl + SM 0 g/L, SM 0.1 g/L, HCl + SM 0.2 g/L, HCl + SM 0.3 g/L, and HCl + SM 0.4 g/L. Additionally, each specimen underwent five repetitions for the respective tests.

2.3. YOLOv4

This study utilizes YOLOv4, an advanced version of the YOLO (You Only Look Once) object detection model, originally introduced by Redmon et al. [43]. YOLOv4, developed by Alexey Bochkovskiy, represents the fourth generation and significantly improves both the speed and accuracy compared to its predecessor, YOLOv3. YOLOv4 consists of three main components, the CSP-Darknet53 backbone, SPP Block, and PANet, and the detection head. The CSP-Darknet53 integrates CSP-DenseNet to enhance the operational speed, addressing the limitations of the previous versions. Various techniques are employed to optimize its performance, the structure of YOLOv4 can be shown in Figure 3. CutMix and Mosaic are the data augmentation methods. CutMix removes objects from training images and replaces them with objects from other images, while Mosaic combines four images into one, reducing the mini-batch size [50]. DropBlock acts as a regularization technique, preventing overfitting by excluding specific regions of the feature map during training. Class Label Smoothing ensures label normalization by allowing label values to range between 0 and 1, enhancing model robustness [51].

The input for this research consists of 3755 image datasets of bubbles. Yolomark was used for labeling and annotation. The output of YOLOv4 is a weights file that records the model’s ability to detect bubbles in a video or image.

Bag-of-Specials (BoSs) techniques in YOLOv4′s detector include the Mish activation function, SPP, SAM, PAN, and DioU-NMS [45]. Mish maintains a non-vanishing gradient, promoting smoother optimization, and serves as a regularizer by bounding negative outputs [52]. SPP ensures consistent output sizes regardless of varying feature map sizes, utilizing multi-filter pooling to preserve important contextual features [53]. SAM conveys spatial-wise information efficiently [54], and PANet transfers information between different levels, improving the localization accuracy [55]. DioU-NMS, the post-processing technique, uses Intersection over Union (IoU) for loss calculation, considering both the overlap and distance between predicted and actual bounding boxes. NMS is applied to remove overlapping bounding boxes, retaining only highly reliable ones [56].

These YOLOv4 components contribute to enhanced performance, efficient computation, accurate localization, and reliable post-processing, making it a powerful tool for object detection in various applications. This study utilizes this advanced model to detect bubbles on the surface of welding metal, demonstrating the effectiveness and versatility of YOLOv4 in real-world scenarios.

In this study, for the bubble detection method, the multi-object-based detection strategy was applied. The dataset was taken through digital camera acquisition and a macro lens. It consists of 3755 images of bubbles, which are divided into 80% for training and 20% for testing.

Figure 4a shows two different conditions for the same welding metal, before and after corrosion. Before corrosion, the metal surface looked relatively homogenous and there were no visible boundaries between the weld metal, heat-affected zone, and base metal. Figure 4b shows an example of a common polarization curve and the timestamp of the video acquisition. The video data were taken for 10 min using a digital camera (DSLR) and a macro lens with 3.2× magnification, which produced a size of 1280 × 720 pixels and 50 fps. Each image has a size ranging from 1 to 2 MB. The size of the image for the data training resulting from the acquisition process is 352 × 352 pixels, with sizes ranging from 100 to 300 KB. For labeling, annotation was accomplished by using YOLOmark [57], labeling in YOLOv4 by marking objects with box-shaped delimiters and adjusting the width and height of the object. The data label format in YOLOv4 has the extension .txt. A data analysis for bubble detection was performed using the images captured within a 6 to 10 min timeframe with a consideration to ensure good image clarity.

Model training and testing employed an AMD Ryzen 7 5800H with 16 GB RAM. The operating system was Windows 11 with a single NVIDIA GeForce RTX 3060 6 GB GPU RAM, which was considered sufficient for training and testing in this study. The setting model configuration of YOLOv4 is shown in

Table 2. Model configuration of YOLOv4.

Parameters	Values	Parameters	Values
Batch	64	Angle	0
Subdivision	32	Learning rate	0.001
Width	352	Burn_in	1000
Height	352	Max_batches	6000
Channels	3	Policy	Steps
Momentum	0.949	Steps	4800; 5400
Decay	0.0005	Scales	0.1:0.1

.

3. Results and Discussion

3.1. Weldment Structures and Hardness

Figure 5 reveals the macro- and microstructure of the specimen’s welded joint. There are three zones of interest, namely, the weld metal (WM), heat-affected zone (HAZ), and base metal (BM), as illustrated in Figure 5a. The weld metal microstructure, as shown in Figure 5b, displays a microstructure characterized by grain boundaries of the acicular, polygonal form of ferrite and perlite, with an average grain size of 21.5 μm. The weld metal’s (Figure 5b) varied microstructure, with large and small grains, stems from the complex thermal cycles in Shielded Metal Arc Welding (SMAW). Rapid heating and cooling due to the welding arc’s high heat cause differing cooling rates across the weld. Slower cooling fosters larger grains in the center, while faster cooling near the edges forms smaller grains. This heterogeneous microstructure results from the SMAW’s localized high-heat input, influenced by the welding parameters and base metal composition. Furthermore, a microstructural inspection of the heat-affected zone, as shown in Figure 5c, is a mostly acicular ferrite structure and has small amounts of perlite with a grain size of 12.8 μm. The base metal zone contains ferrite and perlite phases with an approximation size of the grain of 10.3 μm as shown in Figure 5d. The WM zone acquires the lowest average hardness at 167.1 VHN and then is followed by the HAZ at 173.1 VHN and the BM zone at 192.0 VHN. A Vickers hardness profile of the specimen is presented in Figure 6.

3.2. Potentiodynamic Polarization

Figure 7 shows the potentiodynamic polarization curves of the specimens tested with different concentrations of SM. The vertical axis represents i_corr (current density), which is a measure of the rate of corrosion of a metal or alloy in an electrolyte and expressed in units of amps per square meter (A/cm²). The horizontal axis represents the potential of the electrode in an electrolyte when the potential is varied at a selected rate by applying an electrical source and measured in volts (V) or millivolts (mV) relative to a reference electrode. The curve shift is not easily visible for the three zones as they overlap with each other. The corrosion parameters calculated based on Tafel extrapolation are given in

Table 3. Corrosion parameters calculated from PP curves.

Inhibitor Concentration (g/L)	βa (V/dec)	−βc (V/dec)	Ecorr (V)	icorr (×10−4 A/cm2)	Corrosion Rate (mm/Year)	IE (%)
WM
0	0.122 ± 0.02	0.145 ± 0.03	−0.495 ± 0.01	4.28 ± 0.92	17.78 ± 0.89	0
0.1	0.130 ± 0.02	0.272 ± 0.02	−0.488 ± 0.01	2.72 ± 0.81	11.30 ± 0.57	36.46
0.2	0.124 ± 0.02	0.137 ± 0.03	−0.479 ± 0.01	2.08 ± 0.87	8.63 ± 0.43	51.45
0.3	0.124 ± 0.02	0.146 ± 0.03	−0.483 ± 0.01	2.08 ± 1.02	8.66 ± 0.44	51.29
0.4	0.110 ± 0.02	0.117± 0.03	−0.465 ± 0.01	1.74 ± 0.32	7.21 ± 0.36	59.28
HAZ
0	0.085 ± 0.03	0.095 ± 0.03	−0.493 ± 0.01	1.09 ± 0.05	4.54 ± 0.23	0
0.1	0.053 ± 0.03	0.054 ± 0.03	−0.497 ± 0.01	7.29 ± 0.36	3.03 ± 0.15	33.32
0.2	0.044 ± 0.03	0.043 ± 0.03	−0.495 ± 0.01	5.20 ± 0.26	2.16 ± 0.11	52.38
0.3	0.067 ± 0.03	0.065 ± 0.03	−0.488 ± 0.01	1.02 ± 0.05	4.24 ± 0.21	6.62
0.4	0.085 ± 0.03	0.110 ± 0.03	−0.494 ± 0.01	1.05 ± 0.05	6.84 ± 0.34	4.28
BM
0	0.123 ± 0.02	0.372 ± 0.03	−0.476 ± 0.01	1.05 ± 0.05	4.37 ± 0.22	0
0.1	0.084 ± 0.03	0.256 ± 0.03	−0.469 ± 0.01	6.67 ± 0.33	2.77 ± 0.14	36.56
0.2	0.072 ± 0.02	0.217 ± 0.03	−0.463 ± 0.01	6.59 ± 0.33	2.74 ± 0.14	37.32
0.3	0.090 ± 0.02	0.237 ± 0.03	−0.453 ± 0.01	6.89 ± 0.34	2.86 ± 0.14	34.48
0.4	0.101 ± 0.03	0.283 ± 0.02	−0.443 ± 0.01	7.45 ± 0.37	5.67 ± 0.28	29.02

. SM exhibited an inhibitory effect on the WM zone, as the current density decreased with an increasing concentration. At a concentration of 0.4 g/L, the WM zone showed the lowest corrosion rate of 7.21 mm/year with an inhibition efficiency of 59%. However, at the same concentration, the HAZ indicated an efficiency of 4.28%, with a corrosion rate of 6.84 mm/year. The BM zone had the highest inhibition efficiency of 37% at an extract concentration of 0.2 g/L, resulting in a corrosion rate of 2.74 mm/year.

3.3. Observation of Bubbles Detection and H₂ Evolution

Throughout the duration of the PP test, there was an evolution in the characteristics of the H₂ bubbles, based on the images taken at 6 and 10 min (Figure 8). In the blank solution, larger bubbles were observed, initially adhering to the surface before eventually bursting out. Bubbles continued to form intermittently on the specimen’s surface, indicating a partial cathodic reaction taking place to reduce the H⁺ from the H⁺-rich solution (1 M HCl) [38,58]. Conversely, the specimens placed in the inhibitor solution displayed smaller bubbles that diffused across the surface.

The training process of YOLOv4 in this study consists of several steps as below:

• Dataset Preparation: Gather 3755 images of the bubble dataset with annotated images and the corresponding bounding boxes around the objects of interest, along with the class labels.
• Configuration: Adjust the hyperparameters in the configuration files. Modify the settings, such as the batch size, learning rate, and input image size (Table 2).
• Training Execution: Execute the training process using the prepared dataset. The model learns to detect objects by minimizing detection errors through optimization.
• Loss Calculation: Compute the loss functions (localization, confidence, and classification) based on the differences between the predicted and true bounding boxes, objectness scores, and class predictions.
• Backpropagation and Optimization: Backpropagate the calculated loss to optimize the model’s parameters using optimization techniques, like stochastic gradient descent (SGD) or its variants (Figure 9a).
• Model Evaluation: Periodically assess the model’s performance on a validation dataset using metrics, such as precision, recall, and mAP (mean average precision) (Figure 9b).
• Fine-tuning and Iteration: Fine-tune the model by adjusting the hyperparameters or training it on additional data if needed and iterating until the desired accuracy is achieved.
• Model Deployment: Deploy the trained model for inference on new data or integrate it into applications for real-time object detection tasks (Figure 10).

In order to gain deeper insights from the outcomes, we analyze several calculation results of the detection models trained with the dataset. These results include the mAP50, mAP75, FPS, average IoU, precision, recall, and F1-Score. Precision serves as a metric to assess the accuracy of the model’s predictions by determining the proportion of actual truths correctly identified by the model. Recall measures the ratio of actual truths correctly identified by the model out of the total number of actual truths. The F1-Score is a comprehensive metric that evaluates the balance of the model’s data. It combines precision and recall by calculating their product and dividing it by their sum, multiplied by 2. The IoU, on the other hand, is an indicator that quantifies the extent of the overlap between the bounding box predicted by the object detection model and the actual object. This value is obtained by dividing the overlapped region of the two bounding boxes by the total area covered by the bounding boxes in the image.

In the context of mAP, the term “mAP” stands for mean average precision. AP represents the area under the precision–recall graph. Specifically, mAP is the average value obtained by summing the AP values for each class and dividing it by the total number of classes. A value is usually appended after the mAP to signify the threshold used to determine the reliability of the bounding box when calculating the mAP. The values were recorded at regular intervals, typically every 1000 iterations, during the training process. As depicted in Figure 9a,b, the training has been conducted effectively by reaching a point where the loss has significantly reduced, the accuracy has plateaued, and further improvements are minimal.

Table 4. Detailed testing results.

	mAP	Average IoU	Precision	Recall	F1-Score
[email protected]	97.11%	60.23%	0.77	0.98	0.86
[email protected]	50.48%	43.34%	0.53	0.66	0.59

presents a detailed overview of the test results based on the mAP of the dataset.

The mAP in YOLOv4 serves as an indicator of the bounding box accuracy, which can vary when the model is exposed to new environments or scenarios involving bubbles. However, due to the supervised learning approach employed, the model may struggle to recognize unfamiliar objects [59]. Figure 10 shows that there are a few bubbles that the model fails to identify, and there is a noticeable distinction between the right and left sides of the metal surface. The darker appearance on the right side indicates greater corrosion (after the corrosion process) compared to the relatively less-corroded left side.

The number of bubbles detected in the PP test is given in

Table 5. The number of bubbles detected per unit area (mm²).

Inhibitor Concentration (g/L)	WM	HAZ	BM
0	2.411± 0.121	2.950 ± 0.148	2.025 ± 0.101
0.1	1.032 ± 0.052	1.314 ± 0.065	1.252 ± 0.062
0.2	0.652 ± 0.003	0.612 ± 0.031	1.055 ± 0.053
0.3	0.903 ± 0.045	1.091 ± 0.055	1.837 ± 0.092
0.4	0.353 ± 0.018	1.245 ± 0.062	1.989 ± 0.099

. The SM exhibited distinct effects in the WM, HAZ, and BM zones. Subsequently, this count was further divided according to the specific welding areas, namely, the WM, HAZ, and BM. By dividing the number of bubbles detected by the respective area of each welding zone, the ratio between the bubble count and the area for each specific welding area at a given time is obtained.

The measured volume of the H₂ gas (Figure 2b) that evolved during the PP test (from 0 to 10 min) is given in

Table 6. The measured volume of hydrogen gas that evolved during PP test (×10⁻⁴ mL/mm².s).

Inhibitor Concentration (g/L)	WM	HAZ	BM
0	5.05 ± 0.26	2.44 ± 0.12	2.81 ± 0.14
0.1	3.09 ± 0.16	2.13 ± 0.11	2.27 ± 0.11
0.2	3.06 ± 0.15	1.58 ± 0.08	1.38 ± 0.07
0.3	3.08 ± 0.16	2.15 ± 0.11	2.57 ± 0.13
0.4	2.47 ± 0.12	2.30 ± 0.16	2.63 ± 0.13

. The SM demonstrated a different effect on the weld metal (WM) zone and heat-affected zone (HAZ)–base metal (BM) zone.

Figure 11 shows the comparison between the hydrogen rate and corrosion rate of the specimens tested with different concentrations of SM. The graph shows a similar trend for the H₂ evolution rate and corrosion rate.

3.4. SEM and FTIR Observation

Based on Table 3, the maximum inhibition efficiency of the SM solution for each zone is WM 0.4 g/L, HAZ 0.2 g/L, and BM 0.2 g/L. An immersion test was carried out between the specimens with the SM solution and without the SM solution, which aims to obtain an overview of the surface morphology of each specimen. The surface morphology of the specimens after 5 days of immersion are shown in Figure 12. On the metal surface, several cracks and deposits occurred after being immersed in HCl and HCl + SM. However, the layer of sediment formed in HCl + SM is more even than when immersed in HCl (arrows indicate the areas with minimal sediment). This observation shows how the inhibitory effect of the SM solution restrains the corrosion process.

Figure 13 depicts the FTIR spectra of the SM [60] and all the specimens after 5 days of immersion in 1 M HCl and 1 M HCl + SM. Different bonds appeared in the spectra of the weldment specimens when compared to the original SM spectra. There was a peak of about 1630 cm⁻¹ after immersion in HCl, showing a bending vibration of O-H, and weak absorption around 680 cm⁻¹, indicating a link with Fe (hematite). The intensity of these two peaks decreased after immersion in HCl + SM, indicating the development of hematite groups (stable iron oxide, Fe₂O₃) rather than OH bonds (iron hydroxide).

The examination through an EDS analysis demonstrated a rise in the oxygen content as shown in

Table 7. Detection of elements on the specimen surface by using EDS.

Specimen	Element (wt %)
	C	O	Si	Cl	Mn	Fe
WM in HCl	2.31	1.15	0.55	0.20	0.94	94.85
WM in HCl + SM	2.40	30.07	1.57	0.16	1.03	43.1
HAZ in HCl	2.28	1.16	0.56	0.22	0.75	95.02
HAZ in HCl + SM	2.34	25.62	1.94	0.20	1.45	56.76
BM in HCl	2.68	1.02	1.00	1.88	1.03	92.40
BM in HCl + SM	2.30	21.79	1.99	0.98	1.29	64.25

. This suggests the development of certain oxides that have enveloped the surface of the metal, potentially in the form of Fe(OH)₂ [7].

The hardness test conducted on the three zones of the weldment indicates that the BM zone exhibits the highest level of hardness. This can be attributed to a microstructure characterized by a smaller grain size, which leads to a bigger volume of grain boundaries [17]. The microstructure of the HAZ consists of acicular ferrite and a fine–coarse grain structure. Meanwhile, the primary microstructure of the WM zone is predominantly determined by the grain boundary of ferrite (as depicted in Figure 5). This is associated with a low cooling rate and the process of ferrite formation facilitated by carbon diffusion. Such a microstructure is less effective in inhibiting corrosion compared to a structure with more grain boundaries. Fewer grain boundaries facilitate the diffusion of Cl⁻ ions. The increased volume of the grain boundaries enhances the chemical activities involved in surface corrosion, thereby promoting the formation of a dense and highly stable passive film that provides superior protection [17]. The WM zone exhibits the highest corrosion in the HCl solution due to its larger grain size and lower hardness. It is well established that metals with a lower hardness possess lower corrosion resistance [61]. Furthermore, the presence of precipitates formed during the welding process contributes to the elevated corrosion rate observed in the WM zone [62]. These precipitates often contain non-metallic elements that have the potential to accelerate the corrosion kinetics [62].

During the duration of the PP test, there were observable changes in the properties of the H₂ bubbles (Figure 8 and Figure 11). When analyzing the images taken at the E_corr, it was noted that fewer bubbles were identified on the samples submerged in a solution containing 0.4 g/L of inhibitor, particularly in the WM area. In contrast, a distinct trend was observed in the HAZ and BM regions, where the lowest bubble counts were detected for the specimens immersed in a solution with 0.2 g/L of inhibitor. This variance could be attributed to variations in the microstructural characteristics of the HAZ and BM regions when compared to the WM zone. The coarser grain structure of the WM area does not facilitate the development of a tightly protective layer, resulting in a more active corrosion environment [17] and encouraging the formation of more bubbles.

In the samples without inhibitors, the bubbles tend to exhibit larger sizes compared to those found in the samples containing inhibitors (Figure 8). Furthermore, an analysis of the bubble detection through a deep learning model reveals that the quantity of bubbles is higher in the absence of inhibitors when contrasted with their presence (refer to Table 5). Persistent bubbles appeared intermittently on the surface of the specimen, indicating a partial cathodic reaction to reduce the presence of H⁺ ions in the H⁺-rich solution (1 M HCl) [35,56]. Following the polarization potential surpassing the corrosion potential E_corr point, there was a gradual reduction in bubble formation, eventually leading to the point where no bubbles were visually observed at the last process of the PP test. This observation aligns with the corrosion process transitioning into the anodic reaction phase. The quantified amount of H₂ gas produced during the PP test (as shown in Figure 11a–c) validates the results obtained from the deep learning-based bubble detection. The rate of H₂ evolution exhibits a parallel pattern with the corrosion rate, consequently reflecting the trend in inhibitor effectiveness, as indicated in Table 3. A direct correlation can be observed: the fewer H₂ bubbles detected, the lower the recorded volume of H₂ gas, and the higher the inhibition efficiency (IE). Hence, conducting a quantitative assessment of H₂ bubble properties and H₂ gas volume should yield a strong correlation with IE.

3.5. Proposed Mechanism of Inhibition

The following description provides the ways through which SM inhibits corrosion. Anodic and cathodic reactions take place in acidic solution conditions (pH < 7) as follows:

$$Fe\to {Fe}^{2+}+{2e}^{-}\left(Anode\right)$$

$${2H}^{+}+{2e}^{-}\to H_2\left(Cathode\right)$$

The following equation explains how surface iron atoms can react with hydrogen ions forming corrosion:

$$Fe+{2H}^{+}\to {Fe}^{2+}+H_2$$

Fe²⁺ is unstable and can form a relatively stable FeOOH film after reacting with O₂. Moreover, the FeOOH changes into the stable iron oxidation Fe₂O₃/Fe₃O₄. Because the generated corrosion layer separates the solution from the steel surface, the corrosion rate gradually decreases as the corrosion production builds up [63].

The negatively charged chloride ions (Cl⁻) can interact with positively charged iron ions (Fe²⁺/Fe³⁺) in the solution. It can aggressively erode the corrosion product layer, enhance the capacity of oxygen species to penetrate, and eventually speed up the corrosion of iron:

$${Fe}^{2+}+{2Cl}^{-}\to {FeCl}_2$$

$${Fe}^{3+}+{3Cl}^{-}\to {FeCl}_3$$

When molybdate is added to a 1 M HCl solution, the chloride ions that were previously trapped between the host layers are exchanged for molybdate ions. In a 1 M HCl solution, the components of the passive film [FeMoO₄ and Fe₂(MoO₄)₃] are generated after the molybdate ions from the double-layered hydroxides are released into the solution and inhibit the corrosion in steel. The following reactions arise [64]:

$${Fe}^{2+}+{MoO}_4^{2-}\to {FeMoO}_4$$

$${2Fe}^{3+}+{3MoO}_4^{2-}\to {Fe}_2{\left({MoO}_4\right)}_3$$

The inhibition mechanism of the sodium molybdate solution in HCl to metal weldment relies on the interaction of molybdate ions with the metal surface through adsorption and passivation. Sodium molybdate, acting as a corrosion inhibitor, binds to the iron oxide/molybdate film formed on the metal weldment surface in acidic environments. This film acts as a barrier, hindering aggressive anions like Cl⁻ from reaching the metal surface and preventing metal cations from leaching into the solution [65]. Furthermore, the adsorption of molybdate ions diminishes the corrosion potential and bolsters the corrosion resistance of metal [66]. Schematic of the inhibition mechanism of Sodium Molybdate can be seen in Figure 14. The effectiveness of the sodium molybdate solution in HCl as an inhibitor hinges on multiple factors, including the concentration, inhibitor type and ratio, and prevailing environmental conditions [66,67,68].

4. Conclusions

This study shows how sodium molybdate is used to suppress corrosion, and steel weldments have a different corrosion behavior. This is validated not just through the outcomes of potentiodynamic polarization and hydrogen collection examinations but also via the detection of bubbles employing the YOLOv4 deep learning model.

• With varying efficacy for each weldment zone, sodium molybdate boosts the weldment’s corrosion resistance in 1 M HCl solution. The oxide layer’s stability is increased by the sodium molybdate’s mixed-type inhibition, which also increases the weldment’s corrosion resistance.
• The weld metal’s maximal inhibitory efficiency of the SM solution reached 59% with 0.4 g/L, the heat-affected zone at 52% with 0.2 g/L, and the base metal at 37% with 0.2 g/L of extract.
• The results of training the deep learning model using 3755 bubble images on the mAP50 reached 97.11%, showing that the minimum average bubble detected for the WM was 0.353 /mm² at an SM concentration of 0.4 g/L, while the HAZ was 0.612 /mm² at 0.2 g /L, and the BM was 1.055 /mm² at 0.2 g/L.
• The results from the three conducted tests—potentiodynamic polarization, hydrogen collection examinations, and bubble detection examinations (using deep learning)—display a consistent pattern regarding corrosion behavior. This suggests the potential application of deep learning techniques to analyze corrosion behavior in different scenarios.

This investigation confirms the efficacy of sodium molybdate as a corrosion inhibitor for steel weldments in acidic environments. Achieving optimal corrosion protection involves utilizing different inhibitor doses for distinct weldment zones. Employing machine learning techniques to analyze bubble generation on the weld surface allows for the identification of corrosion behavior.