Advancements in prediction models for corrosion in oil and gas pipelines

2.2.1 Main research institutions

An analysis of inter-country and inter-institutional partnerships can facilitate an understanding of the manner in which pipeline corrosion prediction models are being researched across countries and institutions. Due to the limitations of CiteSpace in analyzing national and institutional data from CNKI, this paper exclusively presents analysis results derived from Web of Science. The maps of country and institution relationships from January 1, 2015 to July 31, 2024 are depicted in Figure 2(a) and (b), respectively. After comprehensive analysis, a total of 80 countries have contributed to the existing literature on this topic during the period from January 1, 2015 to July 31, 2024. Notably, China, Canada, the United States, Australia, England, South Korea, India and Iran emerged as prominent contributors in this field. The institutional collaboration network map reveals that the institutions with the highest publication count include China University of Petroleum, Southwest Petroleum University, Beijing University of Science and Technology, Malaysia National University of Petroleum Technology, and Royal Melbourne Institute of Technology. In a cooperative network, centrality serves as a metric to assess the significance of a node within the network. Based on the graphical representation, it is evident that Southwest Petroleum University and Beijing University of Science and Technology emerge as the most prominent institutions in this institutional cooperative network.

Figure 2:

Map of partnerships: (a) Country, (b) institution.

2.2.2 Keyword analysis

The keyword cluster analysis is shown in Figure 3. Figure 3(a) illustrates that from January 1, 2015 to July 31, 2024, corrosion prediction model research in the field of Web of Science predominantly focuses on keywords such as “corrosion defect”, “mild steel”, “durability”, “elbow”, “stress corrosion cracking”, “atmospheric corrosion”, “ship grounding” and “corrosion rate prediction”. As illustrated in Figure 3(b), the keywords within the CNKI encompass “corrosion prediction”, “neural networks”, “corrosion rate”, “forecasting model”, “corrosion”, “oil and gas pipeline”, “biometry”, and “corrosion of pipeline”.

Figure 3:

Keyword cluster analysis diagram: (a) Web of science, (b) CNKI.

Figure 4(a) and (b) show the time spectra of keyword clustering in the Web of Science and the CNKI from January 1, 2015 to July 31, 2024, respectively. As can be seen through the time mapping, pipeline corrosion has been repeatedly studied over the past decade, in which more studies have been conducted on marine pipelines, with corrosion environments focusing on the atmosphere, soil and marine. Corrosion analysis focuses on corrosion fatigue, stress corrosion cracking, burst pressure and durability, where accelerated corrosion experiments are often used for the above studies. Relevant literature on corrosion prediction has been published since 2015, with corrosion prediction tools such as neural networks, deep learning, and machine learning.

Figure 4:

The time spectra of keyword clustering in (a) web of science and (b) CNKI.

Through the emergent analysis conducted by CiteSpace, we can gain profound insights into the temporal patterns of sudden keyword surges, thereby facilitating a comprehensive understanding of research focal points during each respective time period. The emergent keyword results calculated by CiteSpace for the period from January 1, 2015 to July 31, 2024 are depicted in Figure 5. Figure 5(a) illustrates a significant surge in the number of articles within the Web of Science database, focusing on keywords such as “fatigue life prediction”, “probability of failure”, “sloshing impact load”, and “grounding event” during the period from 2015 to 2016. This observation highlights the emergence of these topics as prominent research areas at that time. Following the year 2020, the primary research focus underwent a shift towards “weathering steel” and “optimization”. As shown in Figure 5(b), the duration of the sudden results obtained from CNKI tends to be longer. The research areas of “pipeline”, “submarine pipeline”, “corrosion” and “neural network” have emerged as the leading directions, while “safety engineering” and “machine learning” have recently gained attention as research hotspots, indicating untapped potential for further exploration.

Figure 5:

Top 7 keywords with the strongest citation bursts in (a) web of science and (b) CNKI.

Through a comprehensive literature analysis study conducted using CiteSpace over the past decade, it is evident that carbon steel-based pipelines have been the primary focus of research, with particular emphasis on investigating their lifespan. The predominant research approach involves conducting accelerated corrosion experiments to observe both metal corrosion fatigue and corrosion behavior. Currently, there is a scarcity of literature on the application of various algorithms for corrosion prediction. Therefore, it is imperative to conduct research on corrosion prediction models. To achieve this, it becomes essential to commence with an exploration of corrosion mechanisms and prediction methods employed in corrosion modeling, followed by a comprehensive review and synthesis of the mechanisms and methods utilized in previous studies.

3.2.1 CO₂ corrosion

CO₂ is a naturally occurring substance. During the extraction of oil and gas, CO₂ is extracted along with other gases, which can cause corrosion of pipelines (Tantawy et al. 2020). Localized pitting, tuberculation, and platform corrosion are prominent manifestations of CO₂ corrosion (Videm and Dugstad 1990). The process of carbon dioxide corrosion is commonly believed to encompass the following stages.

CO₂ dissolves in water and reacts to form carbonic acid:

Carbonic acid in the solution reacts with iron, forming corrosion of iron:

Fe(HCO₃)₂ is unstable at high temperatures and easily breaks down into FeCO₃:

CO₂ corrosion is affected by temperature, CO₂ partial pressure, high chloride concentration, corrosion product film, erosion, among other factors. The aforementioned factors contribute to severe localized corrosion and perforation, thereby posing significant security risks. The prevailing belief is that the primary constituents of the protective film consist of FeCO₃, Fe₃O₄, FeS and oxides derived from alloy elements. The corrosion of CO₂ is influenced by temperature, which in turn affects the chemical reaction and characteristics of the corrosion product film. As a result, different environmental conditions and pipeline materials exhibit varying degrees of corrosion. The specific conditions of the corrosion film under different temperature influences are presented in Table 1. Below 60 °C, De Waard (De Waard and Lotz 1995) proposed an empirical formula to describe the impact of CO₂ corrosion on carbon steel and low-alloy steel.

where V is the corrosion rate, p_CO2 is the partial pressure of CO₂, T is the temperature.

The partial pressure of CO₂ is also a critical factor influencing CO₂ corrosion. It is widely acknowledged that an elevated p_CO2 level leads to an accelerated corrosion rate at a given temperature. The effect of CO₂ partial pressure on CO₂ corrosion was investigated by Lohodny–Sarc (Lohodny-Sarc 1994), revealing distinct corrosion morphologies of metals under varying CO₂ partial pressures, as presented in Table 2. The influence of pH value on CO₂ corrosion primarily manifests in the formation of corrosion products. At a pH level around 10, the predominant corrosion products are Fe(HCO₃)₂ and FeCO₃. Conversely, at a pH range of 4–6, FeCO₃ is formed as the primary corrosion product. Simultaneously, the pH value exerts an influence on the concentration of H⁺ ions within the solution. As the pH level escalates, there is a concomitant reduction in the concentration of H⁺ ions, consequently leading to a decrease in corrosion rate. The presence of oxygen significantly influences the corrosion behavior of carbon steel in CO₂ environments. In the absence of a protective film, oxygen greatly accelerates the corrosion rate; however, when a protective film is already formed, its impact on the corrosion rate becomes negligible (Mcintire et al. 1990). In addition, the corrosion rate is significantly influenced by flow velocity. Specifically, a higher flow velocity corresponds to an increased corrosion rate.

High chloride concentration is also an important influencing factor. Pessu et al. (2020) explored the influence of chloride ion content on carbon dioxide corrosion of carbon steel, and found that the pitting corrosion was most serious when the chloride ion concentration was 10 % and the temperature was 50–80 °C, but there was a chloride threshold when the temperature was below 30 °C, and the chloride ion concentration had little influence on the corrosion rate. Under ambient temperature and pressure conditions, the presence of Cl⁻ reduces the solubility of CO₂ in the medium, thereby slowing down the CO₂ corrosion rate in pipelines. However, under high-temperature and high-pressure environments, the high-pressure conditions significantly enhance the adsorption capacity of Cl⁻, leading to the adsorption of a large amount of Cl⁻ on the surface of carbon steel. The erosive effect of Cl⁻ creates micro-gaps between the corrosion product film and the carbon steel substrate, weakening the bonding force between the corrosion product film and the carbon steel substrate. This process results in the detachment of the corrosion product film, exposing part of the carbon steel substrate directly to the corrosive medium, accelerating the corrosion rate of the substrate, and potentially causing more severe localized corrosion (Chen et al. 2003).

3.2.2 H₂S corrosion

The expansion of oil and gas exploration has led to an increase in the size of high H₂S content wells, thereby introducing significant corrosion risks to pipeline equipment and oilfield facilities. Consequently, the presence of H₂S gas poses both safety hazards and economic losses.

The solubility of H₂S in water surpasses that of CO₂ and O₂, while its corrosion process in aqueous environments can be described as follows (Vakili et al. 2024):

In the context of H₂S corrosion, the formation of a compact iron sulfide film exhibits a certain degree of corrosion inhibition, whereas an insufficiently dense film transforms iron sulfide into the cathode and accelerates steel surface corrosion. The electrochemical corrosion of H₂S aqueous solution also generates hydrogen as a byproduct, which can compromise the structural integrity of the iron-based matrix and result in hydrogen embrittlement. When the H₂S concentration is low, the corrosion products consist of FeS₂, FeS, and a minor proportion of Fe₉S₈. As the H₂S concentration increases, there is an elevated presence of Fe₉S₈ in the corrosion products. The crystal lattice structure of both FeS₂ and FeS exhibits greater compactness compared to that of Fe₉S₈, resulting in superior protective properties for the former.

A.S. Guzenkova et al. (2021) studied the corrosion rate and nature of corrosion in X60 pipeline steel in Samotlor and Usinsk oilfield simulation solution and saturated hydrogen sulfide solution (NACE) medium. It was found that the corrosion rate was the fastest, pH value was the lowest and the number of pitting corrosion was the highest in saturated hydrogen sulfide solution (Table 3).

3.2.3 Coexistence of CO₂ and H₂S

During the transportation of oil and gas, the combined effect of acidic gases such as H₂S and CO₂ can lead to severe corrosion in pipelines (Farhadian et al. 2023; Hu et al. 2017). In the presence of coexistence of CO₂ and H₂S, it is widely accepted that H₂S can expedite CO₂ corrosion through cathodic reactions while mitigating corrosion via FeS precipitation. The predominant influencing factor in the corrosion process is determined by the relative content of these two species. It can be roughly determined who plays a leading role in the corrosion process by p C O 2 p H 2 S , as shown in Table 4.

Under the coexistence conditions, variations in pH value, temperature, and flow rate exhibit distinct influences on the corrosion behavior. When the pH value is excessively low, an inadequately sulfur-containing iron sulfide (such as Fe₉S₈) predominates as the main product, resulting in the formation of a protective film that lacks sufficient protection. Consequently, this phenomenon further accelerates the corrosion rate of steel. The content of FeS₂ increases proportionally as the pH value of the solution gradually rises. Consequently, under high pH conditions, a film layer predominantly composed of FeS₂ is formed, exhibiting a certain level of steel protection.

The impact of temperature on H₂S and CO₂ corrosion is primarily manifested in three distinct aspects (Huang et al. 2023). Firstly, temperature affects the solubility of gases (CO₂ or H₂S) in the medium. An increase in temperature results in a decrease in solubility, thereby impeding the corrosion process. Secondly, elevated temperatures accelerate the rate of each reaction, thereby promoting the corrosion process. Ultimately, an increase in temperature affects the film-forming mechanism of corrosion products, enabling the film to either impede or facilitate corrosion. The study conducted by Qin et al. revealed that, in the case of X65 steel, the corrosion factors in two different environments (steam of the H₂S and CO₂, dissolved brine of the H₂S and CO₂) follow the order of H₂S>>CO₂ > temperature (Qin et al. 2022). The corrosion mechanism of H₂S and CO₂ acting together in two environments is illustrated in Figure 6. In the aqueous solution, ionization of CO₂ and H₂S leads to their combination with Fe²⁺, resulting in corrosion. Conversely, in the gas phase environment, corrosion predominantly occurs at water location.

Figure 6:

Diagrams of H₂S and CO₂ corrosion mechanisms: (a) In solution and (b) in vapour (Qin et al. 2022).

4.1.1 Empirical model

The earliest proposed empirical models primarily included the Norsok M506 model, which predominantly examined the correlation between temperature, total pressure, pipe wall shear stress, pH, and corrosion rate (Zhao et al. 2022). The accuracy of this model in predicting corrosion rates is higher at high temperatures compared to low temperatures. However, it tends to overestimate pitting, localized corrosion, and other forms of uneven corrosion. Additionally, the calculation of pipe wall shear stress is based on the theory of single-phase flow, indicating that this model is more suitable for scenarios involving uniform corrosion and single-phase flow. The specific formula should be applied to the model in various application scenarios. When the temperature is greater than 300 K, the following formula should be employed:

When the temperature is 288 K use the formula:

When the temperature is 278 K use the formula:

where CR_t is corrosion rate, mm/a. K_t is temperature-dependent constant. The f_CO2 parameter can be expressed by the formula (12). And other symbols is defined as formula (13)–(17).

To expand the applicability of empirical models to multiphase flow scenarios, researchers have sequentially proposed the Jepson model for horizontal slug flow (Karunaratne et al. 2017), the Corpos model (Wang et al. 2022) for multiphase flow pipelines, and the Lafayette model (Luo 2023; Wu 2022) for gas-liquid two-phase flow. These findings demonstrate that the precision of empirical model projections is significantly influenced by the suitability of the experimental context and the fidelity of the data. Furthermore, they underscore the challenges associated with adapting empirical models to accommodate changes in environmental conditions and operational procedures.

4.1.2 Semi-empirical model

Semi-empirical model is a combination of empirical formulas and corrosion mechanisms to achieve a corrosion prediction model that is universally applicable to all types of environments. The initial DeWaard75 model was derived by DeWaard et al., who were the first to investigate semi-empirical models and considered the influence of CO₂ partial pressure and temperature based on the carbon steel corrosion mechanism caused by CO₂. However, experimental studies revealed discrepancies between predicted values and actual values. To address this issue, subsequent models proposed by DeWaard incorporated correction factors for each considered factor in order to minimize the disparity between predicted and actual values. Ultimately, the comprehensive DeWaard model accounted for pH, temperature, and CO₂ partial pressure as influential factors, enabling a more accurate prediction of CO₂ corrosion rate on metal surfaces. The most widely used model is the DeWaard95 model, which has the following expression:

where V_r is corrosion rate controlled by reactive activation, mm/a. V_m is corrosion rate controlled by material transfer, mm/a. T is temperature, K. P C O 2 is CO₂ partial pressure, kPa. p H C O 2 is pH value at the same CO₂ partial pressure. U is medium flow rate, m/s. D is pipe diameter, m.

The DeWaard model, being the most renowned semi-empirical model, has served as a fundamental framework for numerous scholars in their endeavors to formulate novel CO₂ corrosion models. The ECE model integrates the IFE ring flow test data with the theoretical foundation of the DeWaard model, while considering the impact of crude oil on corrosion within the model (De Waard and Milliams 1991; De Waard et al. 1995, 2001; Smith et al. 2003). This integration enhances the predictive accuracy of the model, aligning it more closely with real-world pipeline corrosion conditions. Subsequently, BP Company developed the Cassandra model by incorporating the DeWaard model and integrating the film-forming temperature into the framework (Hedges et al. 2000). Considering that a corrosion product film formed at elevated temperatures can confer certain protective properties, this model plays a pivotal role in predicting corrosion protection.

4.1.3 Mechanistic model

The mechanistic model (Nešić 2007) is a system model that integrates thermodynamic and electrochemical analyses, incorporating only a limited number of parameters obtained through rigorous comparisons with experimental data. The mechanistic model not only exhibits a strong reliance on the physical and chemical properties of the system, but also demonstrates robust capabilities for interpolation and extrapolation. However, the existing correction coefficients for mechanistic models often lack sufficient accuracy, making it challenging to ensure the mutual independence of these correction factors. Consequently, this blurs the elucidation of the reaction process within the prediction formula. Therefore, it is of great importance to establish a generalized mechanistic model applicable to corrosion prediction of oil and gas pipelines. Nešić et al. (Nešić 2007) conducted a comprehensive investigation on mechanistic model. The Nešić model (Nesic et al. 1996) incorporates various factors influencing corrosion in actual operating conditions, such as fluid velocity, composition, and temperature. The Nešić model encompasses a wide range of applications, encompassing the dynamic factors of fluids such as velocity and the composition of liquid and gas phases. This comprehensive approach is crucial for accurately predicting changes in metal corrosion rates across diverse operating conditions. Owing to its robustness and extensive applicability, the Nešić model has gained widespread adoption within the oil and gas industry, emerging as an indispensable tool for corrosion prediction in engineering practice.

4.2.1 Power function model with time change

The main form of this model is:

where A is the corrosion loss of a material in the first year at a certain location, g t is the corrosion time, year. n is the protection performance of corrosion products on materials. C _t is the predicted value of corrosion loss at time t.

Due to its reduced number of variables and ease of application, the model was initially preferred; however, during the experiment, it was observed that early corrosion exhibited a more rapid progression, aligning with the power function model. As the duration of corrosion increases, the subsequent corrosion model aligns more closely with a linear model, prompting some scholars to optimize the model by incorporating a power-linear approach (Panchenko and Marshakov 2016). This model is designed to align with the observed reality of corrosion loss. However, following the optimization process, it became evident that certain variables were challenging to ascertain. Furthermore, the exponential function model is more appropriate for the corrosion rate model. In the absence of appropriate cleaning and denoising of the data, the resulting model will exhibit a significant discrepancy between the predicted and actual values.

4.2.2 Grey theory model

The grey theory model is a mathematical framework capable of handling systems characterized by incomplete information and uncertainty. Given the intricate influencing factors and unpredictable effects associated with pipeline corrosion, this model aligns perfectly with its application context. Consequently, scholars have extensively employed this approach in their corrosion research endeavors. In 2007, Ma et al. employed the grey theory to forecast the extent of corrosion on the bottom of storage tanks (Ma and Wang 2007). Wang utilized the GM(1,1) model to forecast the service life of anchor steel under four distinct operating conditions (Wang et al. 2014). Yu et al. established a quality loss prediction model based on grey theory to anticipate metal quality degradation and validated its effectiveness through relative residual error calculations (Yu et al. 2017). Xie and Quan enhanced the smoothness of initial data sequences in order to optimize the GM(1,1) model for predicting the service life of 16Mn steel (Xie and Quan 2020). Liu and Zhang employed an improved GM(1,1) model to simulate corrosion changes in tank bottom plates (Liu and Zhang 2007). The grey theory model was optimized by Jin et al. through background value optimization and translation optimization techniques (Jin et al. 2023). The gray theory model exhibits low data volume requirements, yet demands higher smoothness in the data sequences. Adequate preprocessing of the data is essential when employing gray theory for analysis and prediction.

4.2.3 Markov chain model

The Markov chain is a stochastic approach that facilitates the transition to the target state (Lee et al. 2023), enabling predictions regarding subsequent corrosion conditions based on knowledge of the initial corrosion situation of the material. The utilization of Markov chain models can effectively address the challenges associated with local corrosion and pitting corrosion prediction. The research on corrosion prediction utilizing Markov chain models primarily focuses on investigating pitting states, as these models are well-suited for addressing problems involving interrelated states at different time points, which aligns with the actual situation of pitting phenomena. Andrea Brenna and colleagues employed Markov chains in a predictive model for stainless steel corrosion, utilizing the chemical composition of chlorides as an input variable (Brenna et al. 2017). The algorithm was utilized to calculate the probability of corrosion, enabling accurate assessment of the risk associated with corrosion. Zhang et al. employed Markov chains, inverse Gaussian processes, and gamma processes to simulate the progression of pipeline corrosion (Zhang and Zhou 2014). During the experimental phase, they observed a satisfactory simulation of pipeline corrosion; however, further training with larger sample sizes is necessary before practical application can be considered.

4.3.1 Artificial neural network model

Artificial neural network (ANN) is a common data mining method (Kamrunnahar and Urquidi-Macdonald 2010; Kamrunnahar and Urquidi-Macdonald 2011) (Figure 10). The algorithm possesses the capability to accomplish a nonlinear mapping process between input and output, thereby conferring it with an inherent advantage in elucidating corrosion, which is a prototypical intricate nonlinear phenomenon. Consequently, it currently stands as the most widely employed machine learning technique within the realm of corrosion modeling. Jiang et al. utilized ANN to accurately predict concrete corrosion in sewer systems (Jiang et al. 2016). After a long training period, the corrosion onset time and corrosion rate of the model were found to be very close to those measured in Australian sewers. To investigate the influence of soil heterogeneity on underground pipeline corrosion, Azoor et al. developed and validated ANN using numerical simulations to efficiently generate corrosion profiles with random realizations under various input variable configurations (Azoor et al. 2021). The findings revealed that the spatial variability of saturation significantly affects the size, depth, and occurrence frequency of the largest corrosion pits. The influencing factors of pitting corrosion in Korean steel pipes from 1998 to 2020 were investigated by Kim et al. using ANN (Kim et al. 2023). Their findings revealed the significant roles played by pipe age, soil resistivity, sulfide concentration, and water alkalinity. Given the utilization of ANN in corrosion prediction, scholars have surpassed their reliance on conventional models and are now pursuing more sophisticated and scholarly approaches. Due to the insufficiency of corrosion data and significant variations in corrosion rates across different conditions, traditional models encounter challenges in learning, leading to limited generalization performance. The corrosion rate is influenced by numerous factors, resulting in a highly intricate linear relationship that poses challenges for conventional ANN models to handle. As the inclusion of an increasing number of variables in ANN models becomes prevalent in research, the limited capabilities of traditional ANN for feature selection and extraction may pose a challenge. Consequently, scholars have enhanced the overall model performance through the integration of learning techniques. Shirazi et al. employed artificial neural networks and the imperialist competitive algorithm to successfully forecast the corrosion rate of 3C steel in seawater environments, thus validating the feasibility of this integrated approach (Zadeh Shirazi and Mohammadi 2016).

Figure 10:

Basic neural structure of BP network (Zhang et al. 2021).

4.3.2 Dose-response function model

The ISO 9223-1992 standard, released by the International Organization for Standardization (ISO) in 1992, quantifies the extent of corrosion and corrosion rate across four substrates namely carbon steel, aluminum, copper, and zinc. Subsequently, numerous scholars employed prevailing corrosion theories and mechanisms to establish dose-response functions (DRF) and conducted corresponding investigations. Meanwhile, the ISO has also released the DRF functions pertaining to four categories of substrate materials. The convenience and accuracy of DRF have been acknowledged by corrosion engineers, thereby garnering increased scholarly attention. The DRF function was combined with a competitive algorithm by Panchenko et al. to assess the reliability of predicting first-year corrosion loss of structural metals in corrosive environments for various categories of atmospheric corrosion (Panchenko et al. 2020). The findings demonstrate the relatively robust performance of the DRF function in predicting values. To further assess its reliability, it is recommended to employ the MAPE index. The reliability of two rotation forest models was investigated by Gavryushina et al. in comparison to the DRF function model (Gavryushina et al. 2023). The findings demonstrated that both rotation forest models exhibited superior prediction accuracy and reliability. The limitations of the DRF function in corrosion prediction are evident, necessitating the selection of an appropriate DRF function based on the prevailing environmental conditions (Tables 7 and 8).

4.3.3 Random forest model

Random forest model is a data-driven ensemble learning model that uses bagging resampling technique and random feature subset to generate multiple relatively independent tree models (Figure 11). The final prediction result of RF is obtained by aggregating the predictions from all tree models for a given input sample. Although the single tree model may exhibit instability and susceptibility to overfitting when applied to small sample data, these issues can be effectively mitigated through the combination method of RF. As an integrated algorithm, RF typically possesses a greater number of layers compared to average shallow models, rendering them more adept at handling data characterized by intricate manifold structures. The advantages of RF lie in their ease of implementation, resistance to overfitting, parallelizability, convergence consistency, insensitivity to noise and lack of need for complicated parameter tuning. RF has consistently demonstrated optimal performance across multiple public datasets, making them widely employed in diverse domains. Moreover, RF can also be employed to assess the significance of variables, thereby offering researchers robust support for investigating the underlying mechanisms of the data. In recent years, RF has been employed by scholars in the field of materials and material corrosion due to their inherent advantages. Aghaaminiha et al. employed diverse algorithmic models to forecast the corrosion rate of carbon steel in the presence of corrosion inhibitors (Aghaaminiha et al. 2021). It is found that the random forest algorithm could achieve the most accurate prediction, with the mean square error ranging from 0.005 to 0.093. Pei et al. employed Cu/Fe sensors to monitor the atmospheric corrosion of carbon steel and utilized a random forest model for predicting the atmospheric corrosion rate of carbon steel (Pei et al. 2020). The findings demonstrated that the RF algorithm outperformed both the support vector machine model and the ANN model in terms of accuracy.

Figure 11:

RFR structure (Xinsheng Zhang 2021).

4.3.4 Support vector machine

Support vector machine (SVM) is a type of supervised learning model commonly employed for classification and regression analysis in the field of machine learning (Figure 12). The advantages of this method encompass excellent performance in high-dimensional spaces, the capability to handle small sample sets proficiently, and the ability to effectively address nonlinear problems through the application of kernel functions. SVM can be used to select and extract features that are important for corrosion prediction, as well as to categorize different levels of corrosion states. Jimenez-Come et al. proposed a hybrid SVM-based algorithm for the automatic prediction of pitting behavior in EN 1.4404, aiming to establish a robust model for this purpose (Jiménez‐Come et al. 2018). The study highlighted that the breakdown potential emerges as the primary variable in pitting modeling, while temperature assumes utmost significance in modeling the breakdown potential. The support vector regression-principal component analysis-chaotic particle swarm method proposed by Peng et al. was employed to optimize corrosion prediction models (Peng et al. 2021). The proposed approach is based on the support vector machine model and enhances prediction accuracy through the utilization of chaotic particle swarm optimization. Moreover, it has been observed that this method effectively enhances both the accuracy and interpretability of the model. It has also been argued that the prediction curves of SVM models differ somewhat from the experimental data due to the consistent and smooth input eigenspace transformations of the kernel program (kernel) used in the SVM models (Pagadala et al. 2023).

Figure 12:

Schematic diagram of support vector machine (Xu 2023).