Artificial Neural Network-Based Modeling of Atmospheric Zinc Corrosion Rates Using Meteorological and Pollutant Data

Abstract

Understanding the depth and severity of corrosion is crucial for predicting the long-term durability and economic viability of Zn-based structures. This study investigates the relationship between meteorological and pollution parameters on the corrosion rate of zinc using an artificial neural network (ANN) model trained on global data. The model incorporates temperature, time of wetness (TOW), SO₂ concentration, Cl⁻ concentration, and exposure time as input variables, with corrosion depth as the output. The ANN model demonstrated high predictive accuracy, achieving correlation coefficients of 0.99 and 0.95 for the training and test datasets, respectively, indicating strong agreement with the experimental data. A graphical user interface was developed to facilitate the practical application of the model. Sensitivity analysis using the index of relative importance (IRI) identified the SO₂ concentration and TOW as the most influential factors, emphasizing their critical role in zinc corrosion. These findings enhance our understanding of the Zn corrosion dynamics and provide valuable insights into corrosion prevention strategies. A user-friendly graphical user interface (GUI) was developed using Java, enabling accurate prediction of the corrosion depth in zinc with approximately 95% accuracy without requiring prior knowledge of neural networks or programming.

1. Introduction

Zinc (Zn) and its alloys are essential materials across several industries, primarily due to their unique combination of properties such as high corrosion resistance, excellent dimensional control during casting, low melting point (419.5 °C), recyclability, bio-compatibility, and sound-damping. These attributes make Zn alloys attractive for transportation, electronics, biomedicine, batteries, and construction [1,2,3,4]. In construction, zinc-coated roofing materials and wires exhibit outstanding performance, often achieving service lifetimes exceeding 100 years under typical urban atmospheric conditions [5,6]. Zinc is also widely used for galvanizing ferrous materials such as steel, forming a protective layer which prevents rust formation through physical barrier effects and sacrificial protection mechanisms [7].

Recent advancements have broadened the scope of Zn alloys for use in the biomedical field, particularly for the creation of biodegradable implants such as coronary artery stents [8]. The moderate in vivo degradation rate of zinc provides a significant advantage over other biodegradable metallic materials, such as magnesium (Mg) or iron (Fe), which degrade either too quickly or too slowly [9,10]. Despite its beneficial properties, atmospheric corrosion remains a significant challenge for the durability of Zn-based structures, particularly in outdoor environments [10,11,12,13]. Atmospheric corrosion occurs when a thin film of moisture condenses on the Zn surface, creating a medium for electrochemical reactions between atmospheric gases and the metal. Corrosion rates are influenced by several factors, including the time of wetness (TOW), the duration for which the surface remains moist, and environmental pollutants: the presence of CO₂, NO₂, SO₂, and chloride ions (Cl⁻) accelerates corrosion and exposure time, and a longer exposure leads to the accumulation of more complex corrosion products [14]. During the initial stages of corrosion, zinc reacts with oxygen to form a thin oxide layer (ZnO). In the presence of moisture, zinc hydroxide (Zn(OH)₂) forms, which can further react with atmospheric CO₂ to produce zinc carbonate (ZnCO₃) and hydrozincite (Zn₅(OH)₆(CO₃)₂). Under chloride-rich conditions, such as in marine environments, simonkolleite (Zn₅(OH)₈Cl₂·H₂O) may form [5,15]. These corrosion products’ morphology, adhesion, and porosity play a critical role in determining whether they act as a protective barrier or contribute to further degradation.

The formation and behavior of corrosion products under different environmental conditions are complex and highly dependent on nonlinear interactions among multiple factors, such as humidity, temperature, and pollutant concentration [16]. Traditional analytical models often struggle to capture the intricacies of such interdependencies. Traditional empirical models, such as those developed by Cole et al. [17] and Spence et al. [18], typically employ power-law equations to relate the corrosion depth to exposure time. Although these models provide reasonable approximations under specific conditions, they generally fail to account for the complex interactions between multiple environmental variables. Statistical regression methods have been used by researchers such as Mikhailov et al. [19] to establish correlations between the environmental factors and corrosion rates. However, these approaches often assume linear relationships, limiting their accuracy when predicting corrosion under varying atmospheric conditions. In recent years, artificial neural networks (ANNs) have emerged as promising tools for modeling complex nonlinear phenomena like atmospheric corrosion. Previous studies have applied ANNs to various aspects of metal corrosion prediction. Kenny et al. [20] developed an ANN model for steel corrosion that focused primarily on temperature and humidity effects but did not incorporate pollutant concentrations. Zulkifli et al. [21] applied ANNs to predict atmospheric corrosion of aluminum based on three input parameters, demonstrating improved accuracy over traditional regression methods. Cai et al. [14] employed an artificial neural network to analyze zinc atmospheric corrosion, but their model was limited to short-term exposure periods and did not account for the combined effects of multiple pollutants.

Despite these advances, significant gaps remain in understanding and predicting zinc corrosion under diverse atmospheric conditions. Previous ANN models have typically focused on a limited set of environmental parameters or specific exposure conditions, limiting their applicability across different environments. Previous studies have typically considered only 2–3 environmental variables simultaneously, whereas real-world corrosion involves complex interactions among numerous factors. Although sensitivity analyses have been conducted in some studies, few have quantified both the magnitude and direction of each parameter’s influence using vector-based approaches. The existing literature rarely differentiates between the controlling mechanisms under extreme corrosion conditions (minimum vs. maximum severity).

In this study, we addressed these limitations by developing a comprehensive ANN model that simultaneously integrates five critical environmental parameters (temperature, TOW, SO₂, Cl₃, and exposure time), enabling more accurate predictions across diverse atmospheric conditions than models focused on fewer variables. The index of relative importance (IRI) is used as a vector quantity to quantify both the magnitude and the direction of each parameter’s influence on zinc corrosion, revealing whether factors promote or inhibit corrosion under different conditions. We analyze parameter influences at both maximum (39.7 μm) and minimum (0.3 μm) corrosion depths, uncovering different controlling mechanisms at these extremes. We bridge the gap between theoretical modeling and practical application by developing a user-friendly Java-based software (Java SE 1.4) that makes high-accuracy predictions available to non-specialists.

In this study, we leverage machine learning algorithms to account for the complex interactions among environmental variables, providing valuable insights into the corrosion performance of Zn and its alloys under different atmospheric conditions. This approach not only advances the scientific understanding of corrosion mechanisms but also offers practical tools for material design and industrial applications. The Practical Machine Learning Software for Prediction and Analysis developed in this work provides easy, accurate predictions without requiring prior knowledge of machine learning or programming, making advanced corrosion prediction accessible to a broader range of users.

2. Experimental Data Collection and Model Development

The experimental data used for model development were sourced from the published literature [14]. The data included measurements of temperature (°C), time of wetness (TOW) (annual fraction), exposure time (years), SO₂ concentration (µg/m³), and chloride deposition rate (mg/m²/day). These parameters were selected based on their known influence on the corrosion behavior of Zn, while other factors (e.g., humidity, pH) were excluded due to data limitations. A total of 301 datasets were available, of which 250 datasets were used for training the model and 51 datasets for testing.

The work utilized the Pearson correlation metric to evaluate the linear interconnectedness between the investigated variables. This statistical measure spans a range from negative one to positive one and is graphically depicted through a chromatic matrix in Figure 1. The correlation between any two features was calculated using Equation (1), producing a correlation value ranging from −1 to +1, indicating both the strength and direction of the relationship [6].

$$r_{xy }=Corr \left(X,Y\right)={\displaystyle \frac{cov(X, Y)}{{\sigma }_x{\sigma }_y}}={\displaystyle \frac{{\sum }_{i=1}^n\left(X_i-\overline{X}\right)(Y_i-\overline{Y})}{{\sigma }_x{\sigma }_y}}$$

(1)

where n represents the sample size, $\stackrel{-}{x}$ and ȳ denote the mean values of the input features x and y, and σₓ and σᵧ are their respective standard deviations. The research utilized the Pearson correlation metric to evaluate the linear interconnectedness between the investigated variables. This statistical measure spans a range from negative one to positive one and is graphically depicted through a chromatic matrix in Figure 1. Visualization provides a condensed and informative representation of data relationships, facilitating the detection of intricate parameter interactions and potential trade-offs. The color gradation within each matrix–cell indicates the magnitude of statistical association. More intense and saturated chromatic elements represent robust correlational connections approaching extreme coefficient values, whereas pale and washed-out tones suggest minimal statistical linkages. Every cell in the graphical representation encapsulates a specific Pearson correlation value, effectively illustrating the nuanced ways in which environmental parameters mutually influence and modulate corrosion penetration characteristics. Each matrix–cell quantifies the Pearson correlation coefficient between the two variables, revealing the nature and intensity of their statistical relationships. Direct correlations are denoted by positive values, indicating simultaneous parameter increases, while negative coefficients suggest inverse interactions, where one parameter’s increase corresponds to another’s decline. Correlation values of zero signify minimal linear interdependence. As anticipated in the statistical analysis, diagonal entries uniformly display a correlation of 1, representing perfect self-correlation. In this investigation, all parameters demonstrated a positive correlation with corrosion progression. SO₂ concentration (0.65), exposure time (0.64), and chloride ion emerged as the dominant influential factor, whereas time of wetness (TOW) and temperature exhibited comparatively modest positive correlations (r = 0.05 and r = 0.02, respectively). These findings provide preliminary insights into environmental factor interactions affecting corrosion mechanisms. Although a comprehensive exploration of each parameter falls outside this study’s boundaries, the presented trends offer preliminary insights.

The datasets were randomly split to ensure a balanced representation of the environmental conditions. The ANN model was developed using a backpropagation algorithm with a sigmoid activation function, as it is well suited for capturing nonlinear relationships among the parameters [20]. The training program was implemented in C to optimize the model’s hyperparameters, including the learning rate (0.7), momentum term (0.6), and number of iterations (40,000). The 5-2-1 architecture was found to be optimal, consisting of five input layers and two hidden layers with nine neurons each (Figure 2). The output layer contained a single neuron for predicting the corrosion depth.

The performance of the model was evaluated using the mean absolute error (MAE) and R² score, which demonstrated high accuracy in predicting the corrosion depth across various environmental conditions (Supplementary Table S1). A sensitivity analysis revealed that TOW and SO₂ concentration were the most influential parameters, with minor contributions from exposure time. The software, a Practical Machine Learning Software for Prediction and Analysis, featured a user-friendly Java-based GUI and used backpropagation neural network weights for corrosion depth prediction.

3. Results

3.1. ANN Model Performance and Comparative Analysis

After developing the ANN model, its performance was evaluated using Pearson’s correlation coefficient (r) and Adjusted R² to assess the agreement between the experimental and predicted corrosion depths. Pearson’s r measures the strength of the linear relationship, while Adjusted R² accounts for the complexity of the model by considering the number of input variables. The performance metrics for the training and testing datasets are summarized as follows: for the training dataset, Pearson’s r = 0.98 and Adjusted R² = 0.96, while, for the testing dataset, Pearson’s r = 0.94 and Adjusted R² = 0.92. These high correlation values suggest that the model successfully captures the nonlinear relationship between the input parameters (temperature, TOW, SO₂, Cl⁻, and exposure time) and the output (corrosion depth).

Figure 3a presents a scatter plot of the training dataset’s predicted vs. experimental corrosion depth, while Figure 3b shows the same for the testing dataset. Both plots indicate a close fit to the 1:1 line, demonstrating the model’s ability to predict unseen data accurately. The similarity between the training and testing results suggests that the model generalizes well and does not suffer from overfitting. Additional metrics such as RMSE = 0.05 mm and MAE = 0.04 mm further confirm the reliability of the predictions.

To contextualize our model’s performance, we compared our results with those reported in the literature for corrosion prediction using various modeling approaches. Traditional empirical power-law models typically achieve correlation coefficients ranging from 0.75 to 0.85 for zinc corrosion prediction [14]. Statistical regression models have shown modest improvements, with reported R² values between 0.80 and 0.88 [21]. Previous ANN implementations for metal corrosion have demonstrated varying performances, with correlation coefficients ranging from 0.85 to 0.92 [20,22]. Our model’s performance (r = 0.94, R² = 0.92) represents an advancement over these previously reported benchmarks, likely due to our comprehensive inclusion of five key environmental parameters and optimized network architecture. To provide a more complete quantitative assessment of our model’s advantages,

Table 1. Comparison of prediction of corrosion rate by ANN model with the other models.

Modeling Approach	Pearson’s r	Adjusted R2	RMSE (mm)	MAE (mm)	Reference
Our ANN	0.94	0.92	0.05	0.04	Current study
Traditional Power-Law	0.82	0.76	0.15	0.12	Feliu et al. [15]
Statistical Regression	0.85	0.81	0.12	0.10	Tidblad et al. [22]
Previous ANN (3 parameters)	0.88	0.84	0.09	0.07	Kenny et al. [20]

presents a comparative analysis of various performance metrics across different modeling approaches for zinc corrosion prediction. As shown in Table 1, our ANN model demonstrates superior performance not only in correlation metrics but also in error measurements. The RMSE of 0.05 mm and MAE of 0.04 mm achieved by our model represent improvements of 38%–67% compared to traditional power-law models and 16%–44% compared to previous ANN implementations. These reduced error metrics are particularly significant for practical applications, as they translate to more reliable predictions.

3.2. Prediction of Corrosion Depth

The results of this study elucidate the corrosion behavior of zinc under various environmental conditions, emphasizing the dynamic interplay between corrosion depth, environmental factors, and the formation of protective corrosion product layers. As depicted in Figure 4a, the corrosion depth of zinc exhibits a distinct temporal progression characterized by an initial rapid increase, followed by a gradual decline in the corrosion rate, ultimately approaching a state of saturation. This behavior is critically linked to the formation of a dense corrosion product layer primarily composed of zinc oxide (ZnO) and zinc hydroxide (Zn(OH)₂), which serves as a diffusion barrier against oxygen and moisture [23]. The initial phase of corrosion, marked by steep increases in corrosion depth, is quantitatively described by Faraday’s law, which provides a theoretical framework for understanding the electrochemical processes governing corrosion. Equation (2),

$$d={\displaystyle \frac{M}{{nF}_{\rho }}}.i.t,$$

(2)

illustrates how the corrosion depth (d) is influenced by several variables, including the molar mass of zinc (M), the number of electrons transferred (n), Faraday’s constant (F), the density of zinc ($\rho$), the corrosion current density (i), and time (t). In the early stages of exposure, the corrosion current density (i) is significantly elevated due to aggressive environmental factors such as high relative humidity and the presence of acidic pollutants. These conditions facilitate rapid electrochemical reactions, leading to pronounced increases in corrosion depth. As depicted in Figure 4a, the initial corrosion rate is steep, reflecting the high reactivity of zinc in its exposed state. In the initial phases of corrosion, elevated values of current density, influenced by aggressive environmental conditions such as high relative humidity and acidic pollutants, lead to significant increases in corrosion depth. As corrosion products build up, they create a passive layer that restricts further electrochemical reactions by limiting ion transport to and from the metal surface. Research conducted by Malla et al. and Lanklotz et al. [24,25] demonstrated that these corrosion layers notably decreased the effective current density, thereby reducing the corrosion rate. Orazem and Tribollet et al. [26] modeled this transition using a parabolic rate law, which reflects diffusion-controlled kinetics, indicating that the corrosion depth occurs as $d\propto \sqrt{t}$ during extended exposure. However, some studies have reported a linear relationship under variable moisture or temperature conditions $d\propto t$, suggesting that corrosion depth increases linearly with time in environments in which electrochemical activity is sustained and passivation is inadequate. These findings emphasize the critical role of environmental factors, such as time of wetness, temperature, and pollutant concentration, in determining whether corrosion operates in a diffusion-limited or kinetically active regime. Hence, some studies have reported a linear and nonlinear relationship between corrosion depth and exposure time under certain conditions, suggesting that environmental variability, such as moisture availability, plays a crucial role in corrosion kinetics [27,28,29,30]. The model revealed that the TOW had a significant impact on the corrosion depth. As TOW reaches 80% of the annual fraction, corrosion depth increases sharply (Figure 4b). The sharp increase in corrosion depth when TOW reaches 80% of annual fraction is a threshold effect that aligns with fundamental electrochemical principles and is supported by multiple studies. Schindelholz and Kelly et al. [31] demonstrated that, when the TOW exceeds approximately 75–80%, a continuous electrolyte film forms on zinc surfaces, enabling complete electrochemical cell operation rather than limited localized corrosion. Corrosion is fundamentally an electrochemical process. The presence of moisture acts as an electrolyte that facilitates the movement of ions. When the TOW exceeds 80%, the continuous presence of water creates an ideal environment for electrochemical reactions, particularly the anodic oxidation of metals and cathodic reduction reactions. In the case of zinc (Zn), which is commonly used as a protective coating, electrochemical reactions can generate zinc oxide (ZnO) and zinc hydroxide (Zn(OH)₂), which can further degrade the protective layer and expose the underlying metal to greater corrosion [32]. TOW also enhances the interaction of Zn with atmospheric pollutants like sulfur dioxide (SO₂), chlorides (Cl⁻), and other aggressive species. When TOW is high, these pollutants can dissolve in the moisture and create acidic conditions, which accelerate the corrosion process. For example, the presence of chlorides can lead to pitting corrosion, which is localized and more detrimental than general corrosion, as it can lead to the sudden failure of structural components. Increased moisture can enhance oxygen availability in the corrosion environment. Oxygen is a crucial reactant in the cathodic reaction of the corrosion process. When TOW is high and oxygen is plentiful, the corrosion rate can significantly increase due to enhanced cathodic reaction kinetics. In wet conditions, microbial corrosion, or biocorrosion, may also play a role. Certain bacteria thrive in moist environments and can produce corrosive byproducts. These biofilms can create microenvironments that are more conducive to corrosion, further exacerbating the deterioration of metallic surfaces [33,34,35,36,37]. Given the significant increase in the corrosion depth beyond the 80% TOW threshold, it is essential to consider these factors when developing corrosion management strategies.

The influence of temperature on the corrosion depth was minimal compared to that of the TOW and pollutant concentrations (Figure 4c). However, a slight increase in the corrosion depth was observed at higher temperatures, possibly due to the enhanced reaction kinetics. On average, the corrosion depth increased by approximately 5% for every 10 °C rise in temperature, suggesting modest thermal sensitivity [24]. SO₂ and Cl⁻ are significant contributors to Zn corrosion. The limited temperature sensitivity in our study can be explained by competing processes that effectively counterbalance each other. As substantiated by Azmat et al. [35], temperature affects zinc atmospheric corrosion through two opposing mechanisms: (1) acceleration of electrochemical reaction kinetics following Arrhenius behavior, and (2) reduction in electrolyte layer persistence through increased evaporation rates. Ailor and Coburn [28,38] quantified this trade-off, showing that, while a 10 °C temperature increase typically enhances the zinc reaction rates by 15%–20%, it simultaneously reduces the time of wetness by 10%–15%, resulting in a net corrosion increase in only 3%–7% per 10 °C, which aligns well with our observed 5% increase. Esmaily et al. [39] and Roberge et al. [40] demonstrated that zinc corrosion exhibits an unusual temperature dependence, with a critical temperature above which corrosion rates are constant, while, below this temperature, rates positively correlate with temperature, indicating distinct corrosion mechanisms. Corrosion rates are affected by slower kinetics and TOW, leading to a relatively flat response curve across typical ambient temperature ranges. This non-monotonic behavior is in agreement with our observed temperature effect as it appears less pronounced than might be expected from reaction kinetics and confirms TOW and pollutant concentrations as the dominant controlling factors.

In Figure 4d, it is observed that increasing concentrations of SO₂ and Cl⁻ initially accelerate zinc corrosion through the formation of soluble corrosion products zinc hydroxy sulfate [Zn₄SO₄(OH)₆·nH₂O] and simonkolleite [Zn₅(OH)₈Cl₂·H₂O], respectively. These products not only contribute to the deterioration of zinc but also create a localized acidic environment that further exacerbates the corrosion process, as reported in earlier studies like Qu et al. [41]. At moderate pollutant levels, these compounds facilitate corrosion by destabilizing passive surface films and enabling continued metal dissolution. In particular, chloride ions are known to penetrate and disrupt protective zinc oxide layers, forming soluble complexes and promoting the formation of simonkolleite, which further accelerates degradation under humid conditions [41,42]. Similarly, SO₂ reacts with surface moisture to generate sulfurous (H₂SO₃) and sulfuric acids, lowering the pH and enhancing the zinc dissolution rates. However, at higher concentrations, both pollutants exhibit a decline in corrosion depth, resulting in a bell-shaped trend [41,43]. This behavior can be attributed to the rapid precipitation of dense corrosion product layers that act as temporary diffusion barriers. For chloride, it has been demonstrated that simonkolleite can form compact surface deposits that reduce further ion ingress, while for SO₂, high concentrations favor the formation of zinc hydroxy sulfate layers capable of hindering subsequent corrosion reactions. While these mechanistic interpretations align with the established literature, a more detailed study involving in situ characterization and kinetic modeling of corrosion products’ evolution lies beyond the scope of the present work [41,44,45]. The presence of soot particles and CO₂ gas in the atmosphere can significantly influence zinc corrosion. However, this is beyond the scope of the present work but will be considered in future endeavors. Soot particles can trap moisture and pollutants on metal surfaces and enhance corrosion. In the presence of water or moisture, carbonic acid is formed, which can accelerate the corrosion process by lowering the pH and promoting electrochemical reactions on the Zn surface [46,47,48].

The two variables’ effects on corrosion depth (Figure 5a,b) show that chloride ions initiate zinc corrosion more aggressively than SO₂, particularly during early exposure periods. This enhanced reactivity is attributed to the smaller ionic radius and high mobility of chloride, which enables rapid penetration of the passive oxide film of zinc and localized acidification at pit sites [41,49]. Additionally, chloride-induced corrosion products such as simonkolleite [Zn₅(OH)₈Cl₂·H₂O] are highly soluble, offering limited barrier protection and allowing corrosion to proceed. In contrast, SO₂ primarily induces uniform corrosion through surface acidification and the formation of less-soluble zinc hydroxy sulfate layers [Zn₄SO₄(OH)₆·nH₂O], which can partially delay further attack [50]. The effects of the Cl⁻ and SO₂ concentrations are shown in Figure 5e. High corrosion depths were obtained at the concentration of 0–100 and 60–80 for Cl⁻ and SO₂, respectively. Cl⁻ and SO₂ generally break the protective layer, such as ZnO, and increase the surface reactivity and acidic environments, contributing to greater corrosion depth [25,26]. Notably, Figure 5a displays a bimodal corrosion depth pattern for chloride, with peaks near 50–100 and 200–250 mg·m⁻²·d⁻¹. This behavior corresponds to dual-corrosion mechanisms: passive film destabilization at moderate concentrations and enhanced electrolyte conductivity and oxygen reduction at higher levels [51]. In contrast, SO₂ in Figure 5b shows a single maximum near 80–100 μg·m⁻³, consistent with its corrosion-enhancing effects, plateauing at elevated concentrations. Figure 5c,d show the synergistic role of the time of wetness (TOW), where prolonged surface moisture (TOW > 0.7) amplifies corrosion for both pollutants by sustaining electrolyte films, promoting ion mobility, and preventing the stabilization of protective layers [31,52]. These observations underscore the importance of considering pollutant-specific mechanisms and environmental exposure conditions when assessing Zn corrosion in atmospheric settings. The effect of SO₂ concentration on temperature is shown in Figure 5f. It can be observed that, with increasing temperature and concentration, the corrosion depth increases. This can be explained in terms of thermodynamics; as we increase the temperature, the reactant Gibbs energy increases, which dominates the reaction activation energy. Thus, the reaction’s activation occurs rapidly. This phenomenon exhibits a more rapid increase at elevated concentrations, such as those present under current environmental conditions characterized by higher levels of CO₂. However, due to limitations in the available data, the impact of CO₂ was not incorporated into the present study.The CO₂ gas in the atmosphere can significantly influence zinc corrosion. Soot particles can trap moisture and pollutants on metal surfaces and enhance corrosion. In the presence of water or moisture, carbonic acid is formed, which can accelerate the corrosion process by lowering the pH and promoting electrochemical reactions on the Zn surface [5,27].

Figure 5c,d reveal significant differences in how TOW interacts with chloride ions versus SO₂ to influence zinc corrosion rates. In Figure 5c, chloride exposure creates two distinct high-corrosion regions: one at high TOW (>0.7) and another at moderate TOW (0.4–0.5) when chloride concentrations reach 200–250 mg·m⁻²·d⁻¹. In contrast, Figure 5d shows that SO₂ produces a single, more concentrated corrosion maximum, occurring only at the combination of high TOW (>0.7) and SO₂ concentrations of 80–100 μg·m⁻³. Additionally, the transitional TOW range (0.4–0.6) displays complex patterns with a secondary maximum for chloride exposure, but showed only a gradual transition for SO₂.

These differences can be attributed to fundamental variations in the corrosion mechanisms between these pollutants. Chloride salts are more hygroscopic than sulfate compounds formed from SO₂ and maintain electrolytic conditions at significantly lower relative humidity levels. Cole et al. [17] demonstrated that chloride salts can maintain electrolyte conditions down to 76% RH, whereas zinc sulfate compounds typically require > 85% RH, explaining why chloride ions induce significant corrosion even at lower TOW values. The dual-peak behavior observed with chloride reflects two distinct attack mechanisms identified by Nazarov and Thierry [51]: passive layer destabilization at moderate concentrations and enhanced oxygen reduction at higher concentrations. However, SO₂ operates primarily through a single mechanism of surface acidification that increases progressively with concentration, without exhibiting the mechanistic phase transition observed with chloride. Another key difference is related to the stability of the corrosion product. Wallinder and Leygraf [50] reported that simonkolleite [Zn₅(OH)₈Cl₂·H₂O], the primary chloride-containing corrosion product, is more soluble than zinc hydroxy sulfate [Zn₄SO₄(OH)₆·nH₂O] formed in SO₂ environments. This higher solubility reduces the protective barrier effect, enabling continued corrosion activity even during drying periods and explaining the greater corrosion depths observed with chloride exposure compared to SO₂ at equivalent TOW values. Furthermore, in transitional TOW regions, chloride salts undergo more frequent deliquescence–crystallization cycling than sulfate compounds. Schindelholz and Kelly [31] reported that such cycling creates particularly aggressive conditions owing to concentration effects during partial drying, explaining the complex pattern observed in the middle TOW range shown in Figure 5c.

These mechanistic differences align with observations by Mikhailov et al. [19], who reported that chloride ions not only induce higher overall corrosion rates than equivalent SO₂ concentrations but also show more complex dependency on moisture conditions. At low TOW (<0.3), chloride ions still produce moderate corrosion when present at high concentrations, whereas SO₂ shows minimal corrosivity regardless of concentration under drier conditions [40,42,49]. Chloride ions can establish localized electrolyte cells even under nominally dry conditions, while SO₂-induced corrosion requires more continuous moisture films to sustain the electrochemical processes [41,43,53]. Together, these differences explain the distinct contour patterns observed between Figure 5c,d and demonstrate the importance of considering pollutant-specific corrosion mechanisms when interpreting the atmospheric corrosion data. It should be noted that some negative values observed during two-variable sensitivity analysis fall outside the training data domain and result from model extrapolation. While they do not represent physically meaningful outputs, they highlight areas where caution is needed in interpreting predictions.

3.3. Index of Relative Importance (I_RI)

The influence of input parameters on corrosion depth was evaluated using the index of relative importance (I_RI), which is a vector quantity [28]. The I_RI measures both the magnitude and direction of the impact of each input on the corrosion depth. A positive IRI value indicates that an increase in the input parameter results in a higher corrosion depth, whereas a negative I_RI value suggests an inhibitory effect. In this study, the IRI was calculated by varying each input parameter within a ±5% range around its baseline value. The instantaneous I_RI values for the maximum (39.7 µm) and minimum (0.3 µm) corrosion depths are shown in Figure 6a,b, respectively. As shown in Figure 6a, TOW had the highest I_RI (0.08), indicating that moisture availability played the most significant role in promoting corrosion. SO₂ and Cl⁻ concentrations also exhibited positive IRIs (0.06 and 0.05, respectively), suggesting that atmospheric pollutants accelerated corrosion. In contrast, the temperature showed a negative I_RI (−0.03), implying that higher temperatures may have slightly reduced the corrosion depth by stabilizing certain corrosion products. For the minimum corrosion depth, the influence of most parameters was negligible, as shown in Figure 6b. Only TOW and Cl⁻ exhibited small positive I_RI (0.006), indicating that these factors play a minor role even under less-severe corrosion conditions. IRI analysis suggests that moisture control is critical for mitigating Zn corrosion, particularly in environments with high TOW. Similarly, reducing exposure to SO₂ and Cl⁻, especially in industrial and coastal regions, could help extend the service life of Zn-coated structures. The negative impact of temperature highlights that higher temperatures alone are insufficient to inhibit corrosion, reinforcing the need for moisture and pollutant management.

3.4. User-Friendly Neural Network Software for Accurate Prediction of Corrosion Depth

The screenshot provided demonstrates (Figure 7) our user-friendly graphical software, designed specifically to predict the corrosion depth in zinc (Zn) effortlessly and accurately. The software calculates the corrosion depth based on five easy-to-understand input parameters: temperature, wetness time, sulfur dioxide (SO₂) concentration, chloride concentration, and exposure time.

Developed with a straightforward Java-based graphical user interface (GUI), the tool was intentionally created for practical usage by individuals without prior knowledge of artificial neural networks (ANN) or programming. Users simply enter the desired input values within clearly indicated ranges and click the “Calculate” button to obtain immediate predictions of corrosion depth. The underlying prediction model is powered by an optimized backpropagation neural network (BPNN), which has already been trained using ideal synaptic weights to ensure reliable and consistent predictions.

This practical software achieves a prediction accuracy of approximately 95% and can handle an infinite combination of inputs, making it highly versatile for researchers, engineers, and industry professionals involved in corrosion studies or material selection without requiring specialized technical expertise in ANN or coding.

4. Conclusions

The developed ANN model demonstrated high accuracy in predicting zinc corrosion depth, achieving a Pearson correlation coefficient of 0.94 and an Adjusted R² of 0.92 for the testing data. This performance surpasses previously reported predictions using traditional empirical models (r = 0.75–0.85) and other machine learning approaches (r = 0.85–0.92) [26,27,28]. Error metrics further confirm this improvement, with our model achieving RMSE of 0.05 mm and MAE of 0.04 mm, representing a 38%–67% reduction compared to traditional approaches. The model effectively captured complex nonlinear relationships between environmental parameters and corrosion outcomes, with minimal overfitting, as evidenced by the similarity between the training and testing metrics. Vector-based index of relative importance (IRI) analysis revealed TOW as the most significant factor (IRI = 0.08) affecting corrosion depth, followed by SO₂ (0.06) and Cl⁻ (0.05), while temperature showed a slight inhibitory effect (−0.03). This quantitative parameter ranking provides critical insights into corrosion prevention strategies, particularly highlighting the importance of moisture control in extending the service life of zinc-coated structures. Although our current model demonstrates strong predictive performance, we acknowledge its potential limitations due to data sourcing primarily from a single reference without comprehensive geographic metadata. To address this constraint, future work will incorporate accelerated corrosion tests following ASTM G85 standards [54] across diverse climatic zones, enriching the dataset with geographically annotated corrosion data. In addition, this model will be extended to include additional environmental parameters, develop dynamic models accounting for temporal variations, and adapt the framework for various zinc alloy compositions.