Predictive Corrosion Degradation Modelling of Oil Country Tubular Goods Using Physics-Informed Machine Learning and Digital Twin Technologies

The oil and gas industry must deal with more complicated and demanding conditions because it is looking for deeper reservoirs, and the world requires more energy to sustain its growth. Sour service environments are some of the most difficult to work in because high temperatures, high pressures, and corrosive substances such as hydrogen sulfide (H₂S), carbon dioxide (CO₂), and chlorides work together to accelerate the breakdown of Oil Country Tubular Goods (OCTG). These are important steel components that maintain wellbore stability during drilling and production. This is because there are numerous ways in which things can break down, and they play off each other. These include uniform corrosion, localized pitting, stress corrosion cracking (SCC), sulfide stress cracking (SSC), hydrogen-induced cracking (HIC), and corrosion-fatigue coupling. Such processes compromise the structural integrity of wells and pipelines and predispose them to disastrous failures that can have both environmental and financial consequences [1, 2].

The primary approaches to traditional maintenance and qualification work consist of standardized laboratory tests and empirical correlations based on the number of years of industrial experience. These techniques have provided valuable data on the behavior of materials; however, because they are grounded in actual data, they are not always able to predict how materials will behave under varying conditions of service. Traditional testing or models at rest cannot wholly show alterations in the sour environment in terms of temperature, fluid chemistry, and stress. Consequently, maintenance approaches based on these techniques tend to either be overly conservative and costly, or fail to eliminate the possibility of failure too early [3, 4]. To overcome these issues, predictive models that integrate experimental data, physical theory, and real-time data of operations are needed to ensure a better demonstration of the degradation evolution.

Recently, through the integration of research in material science and artificial intelligence, a novel computational tool called physics-informed machine learning has emerged with impressive applications in the analysis of complex engineering systems. In contrast to purely data-driven approaches, PIML models involve the direct embedding of constitutive equations, conservation laws, or governing principles into the learning framework to ensure physically consistent solutions with high interpretability, even with limited data. PIML differs from traditional data-driven models in that it explicitly uses governing equations and constitutive relationships during the learning process. This means that the model can be comprehended and obeys the simple laws of physics (even in cases where little data are available) owing to this integration [5, 6]. Such models can be used to describe fatigue crack propagation and corrosion kinetics using physically based constraints, such as the law of Paris and Arrhenius-type relations, in the context of OCTG reliability. PIML improves the prediction accuracy and helps reduce the incidence of nonphysical results by combining these equations into a learning system. The combination of real-world data obtained in the real world and mechanistic knowledge enables the development of both experimental validation and predictive analytics, which is its hybrid quality.

Digital twin technology also plays a significant role in data-driven reliability management. A digital twin refers to a model of a physical object that is computer-generated and evolves over time as it is continuously updated with field measurements obtained through sensors. In OCTG systems, digital twins integrate data on pressure, temperature, and chemical composition with predictive algorithms to determine the state of the material, its degradation, and its life expectancy. Constant monitoring and simulation enable engineers to enhance inspection and maintenance schedules [7, 8] to provide constant monitoring and risk estimation rather than adherence to defined timeframes. The addition of PIML to digital twin architectures produces models that can be modified and self-healed with shifts in operations. This enables the maintenance strategies to become more dependable and economical.

The combination of data analytics, computer modelling, and materials science is transforming energy sector operations. It has been shown that hybrid models that combine physical laws and machine learning algorithms offer greater adaptability, scalability, and interpretability. Such models dynamically adapt to service conditions, forming closed feedback loops that lead to improved safety and durability [9, 10]. Moreover, complex simulation procedures, such as Deep Fluids and hybrid damage identification algorithms, have demonstrated how deep neural networks can accurately recreate both thermomechanical and electrochemical processes with a high level of computational efficiency [11, 12]. Methods are necessary to record the mechanical and chemical interdependent degradation phenomena that govern the behavior of OCTGs.

Despite these improvements, research gaps remain significant. The majority of studies emphasize either PIML or digital twin technologies independently, which makes the approaches to this fragmented and thus fails to merge the experimental data on mechanical, electrochemical, and microstructural processes into a unified framework. In addition, many of these studies are limited to laboratory experiments or theoretical studies, which limits their extrapolation to field-scale operations, where loads and chemical compositions are variable [13, 14]. To overcome these limitations, the development of holistic hybrid systems that integrate continuous field data with physical degradation models is critical to making them adaptable and interpretable.

These gaps were directly filled in this study with a unified physics-constrained artificial intelligence and digital twin framework for predictive degradation assessment and lifespan optimization in OCTG systems operating under extreme service conditions. It is a framework that unites experimental characterization, machine learning, and physical modelling into one system capable of illustrating the breakdown of things in the real world. Constitutive relationships are involved in the learning process, which defines the distribution of fatigue cracks and the corrosion process. This ensures that model predictions are always physically consistent and reliable, irrespective of the conditions [15, 16]. The digital twin element implements a predictive structure by incorporating real-time field measurements to continuously refresh the degradation and remaining life projections to enable proactive maintenance and inspection scheduling based on risks.

The proposed study contributes to the enhancement of predictive reliability engineering by unifying physical principles, artificial intelligence, and real-time data analytics in a unified system. PIML and digital twin technologies represent a significant shift in asset management towards a proactive approach rather than a reactive one. This approach decreases unexpected downtimes, minimizes maintenance expenses, and enhances safety and environmental standards [17]. Along with OCTG systems, the general concept of this hybrid structure can be applied to other applications, including aerospace, transport, and biomedical engineering, in which materials must perform in harsh environments. This research paper is part of a worldwide movement in intelligent, sustainable, and resilient engineering systems that rely on physics-informed digital transformation, which combines real-world data with computer-based forecasts.

3.1 Mechanical performance

The average values of the data collected from the three samples under similar loading conditions for all three alloys are presented in Table 1 and Figure 1. The data show improved values of both strength and ductility as a result of advancements in the alloy composition and microstructure.

Table 1. The tensile properties of all materials in this study

API X70 carbon steel has moderate strength and low ductility, which are typical of ferritic-pearlitic microalloyed steels commonly used as standard OCTG casing and line pipe products [13, 18]. The yield strength of 482 MPa and elongation of 12.4 percent are in line with the reported data on the yield strength of hydrogen-exposed microalloyed steels, where plasticity is decreased by embrittlement and segregation of carbide at the grain boundaries [20]. These microstructural constraints are commonly associated with premature fatigue crack propagation during sour service, which is also the case for comparable X-series steels [27].

Figure 1. Comparison of yield strength and ultimate tensile strength (UTS) for API X70, 22Cr duplex stainless steel, and Inconel 625

22Cr duplex stainless steel, on the other hand, has much better strength and ductility. The austenitic-ferritic dual structure improves the resistance to dislocation motion and strain accommodation, which leads to a yield strength of 552 MPa, ultimate strength of 748 MPa, and elongation of 25.7. These findings are consistent with those of Kılınç et al. [30] and Sun et al. [27], who demonstrated that duplex steels can maintain their toughness and mechanical integrity during simultaneous exposure to CO₂/H₂S because of their balanced phase distribution.

The large elongation is also evidence of the effective trapping of hydrogen and deflection of cracks at the ferrite-austenite boundaries, enhancing their resistance to localized stress corrosion [28].

Inconel 625 exhibited the best performance, with a yield strength of 603 MPa, ultimate tensile strength of 903 MPa, and elongation of 39.6 percent. The cause of this behavior is the solid-solution strengthening of Mo and Nb, coupled with dislocated γ′ and γ″ precipitates, which increase the creep and fatigue strength [29, 31]. Solovyeva et al. [29] reported that thermal stability and plasticity of Inconel 625 render it highly appropriate in high-temperature and high-H 2 S conditions in which typical steels tend to lose their ductility quickly. Moreover, its high microstructural refinement-fatigue endurance correlation promotes previous computational forecasts of digital twins and physics-informed models [7, 21, 22, 31, 33].

The strengthening and ductility improvement between API X70 and 22Cr Duplex and Inconel 625 prove that the enhancement is possible because of the design of alloys and phase control.

This confirms that high-nickel and duplex systems are more applicable to sour-service OCTG components that require high load capability and long-term dimensional stability [11, 21, 22].

3.2 Electrochemical behavior

The electrochemical data presented in Table 2 and Figure 2 were derived from Tafel polarization and EIS experiments performed in a 3.5% NaCl solution saturated with CO₂ (0.1 MPa) and H₂S (0.05MPa) at 80℃. As shown in Figures 3 and 4, there was a strong composition dependence of the corrosion rate and passive layer lifetime.

Table 2. Electrochemical parameters in CO₂/H₂S-saturated brine at 80℃

Figure 2. Comparison of corrosion current density (icorr) and corrosion rate for the investigated materials in CO₂/H₂S-saturated brine at 80℃

Figure 3. Tafel polarisation curves of API X70, 22Cr duplex steel and Inconel 625

API X70 carbon steel recorded the highest corrosion current density (30.1 µA/cm²) and lowest charge-transfer resistance (0.93 kΩ cm²), resulting in a corrosion rate of approximately 0.349 mm/year. This effect is normal in low-alloy ferritic steels in sour environments, where the presence of CO₂ and H₂S destabilizes the FeCO₃/FeS protective layers, resulting in localized film breakdown and accelerated anodic dissolution [27, 29]. The interaction between CO₂ and H₂S enhances the cathodic hydrogen evolution reaction, increasing surface acidity and causing pit nucleation [27, 29].

Figure 4. Nyquist impedance plots of API X70, 22Cr duplex steel and Inconel 625

Nyquist plots typically have a small semicircle diameter of X70, indicating that the main control of corrosion is charge transfer rather than diffusion, as observed by Solovyeva et al. [29].

The 22Cr duplex stainless steel exhibited a significantly low corrosion current density (7.0 µA/cm²) and high Rct (4.06 kΩ cm²), which was equivalent to a corrosion rate of 0.066 mm/year.

This is because it forms a small Cr₂O₃–FeCr₂O ₃ passive layer, which is resistant to the CO₂-H₂S combination [27, 30]. Kılınç et al. [30] stated that duplex stainless steels form a duplex oxide/sulfide coating that keeps the structure intact despite the long exposure period and drastically minimizes hydrogen ingress. In addition, the ferrite–austenite interface provides a network of hydrogen traps, reduces embrittlement, and controls the anodic dissolution rate [28, 33].

Inconel 625 exhibited the best corrosion resistance, lowest corrosion current density (1.1 µA/cm²), and maximum Rct (11.85 kΩ cm²); consequently, an exceedingly low corrosion rate of 0.011 mm/year was obtained. This is linked to the development of a thick NiO-Cr₂O₃-MoO₃ passive layer, which offers better resistance to both chloride and sulfide cation permeabilities [29, 31]. According to Wu et al. [33] and Meza et al. [34], Ni-based superalloys exhibit diffusion-controlled reactions to impedance under the conditions of H₂S/CO₂, which implies that corrosion is constrained by slow ionic conduction through the protective film rather than by charge-transfer reactions.

The steep phase angle in the Bode diagrams and the large diameter of the semicircle in the Nyquist plots attest to the fact that the signature of stable passive behavior is in line with the predictions of the long-term behavior of the digital twin in sour service environments [31, 34].

In general, Inconel 625 exhibits a higher resistance to corrosion than 22Cr Duplex SS and API X70, which is negatively proportional to icorr and directly proportional to Rct. These findings clearly show that more protective oxide-sulfide layers are formed in the presence of higher Cr, Ni, and Mo alloys, which significantly decreases the rate of corrosion in CO₂/H₂S settings [27-30].

Moreover, the parameters obtained in the polarization and EIS analyses (icorr, Rct, and corrosion rate) are crucial input parameters in the next physics-informed machine learning and digital twin models applied in predictive degradation modelling and reliability measurements [21, 31, 33, 34].

3.3 Microstructural analysis

High-resolution SEM enabled microstructural analysis to provide the necessary information on the fracture mechanisms that dominate the behavior of the three alloys examined under sour service conditions. As shown in Figure 5, SEM observations were used to complete the mechanical and electrochemical observations, showing how the different fracture morphologies, including brittle cleavage in API X70, mixed-mode characteristics in 22Cr duplex steel, and fully ductile microvoid coalescence in Inconel 625, were determined by the underlying deformation modes and environmental susceptibility of each material.

Figure 5. Fractographic features: (a) brittle cleavage in API X70, (b) mixed quasi-cleavage and ductile dimples in 22Cr duplex steel, and (c) fully ductile microvoid coalescence in Inconel 625

For API X70 carbon steel, the SEM fractograph showed a strong cleavage facet with secondary cleavages, representing a predominantly brittle and transgranular fracture process. The angular fracture planes and scanty plastic relief indicate that hydrogen-assisted cracking and localized anodic dissolution work together to propagate cracks under stress in a brittle manner. These fracture features can be attributed to the fact that CO₂/H₂S co-exposure disrupts the protective FeCO₃/FeS coating and increases crack propagation in ferritic steels [27, 28].

The 22Cr duplex stainless steel exhibited a mixed-mode fracture morphology with quasi-cleavage facets and ductile microdimples. This mixture results from the bipolarity of duplex alloys, with the ferritic component being more prone to fracture and the austenitic component being more prone to plastic deformation. This phase-selective fracture behavior increases crack bending, energy capture, and toughness, as observed in duplex alloys exposed to hostile environments [30, 31].

In contrast, Inconel 625 exhibited a fully ductile fracture morphology typified by evenly dispersed hemispherical dimples and definite signs of microvoid coalescence. The lack of brittle characteristics or secondary cracking emphasizes the remarkable hydrogen uptake and localized corrosion resistance of the alloy. This ductile nature is consistent with the long-established effect of γ′/γ″ strengthening precipitates in Ni-based superalloys, which facilitates homogeneous deformation and slower crack growth, even under harsh sour conditions [33, 34].

As shown by the SEM evidence, brittle cleavage (API X70), mixed-mode behavior (22Cr duplex stainless steel), and full ductile microvoid coalescence (Inconel 625) were distinctly connected and directly corresponded to the trends in the mechanical, electrochemical, and degradation properties of the three materials.

3.4 Physics-informed AI model performance

Table 3 shows that the evaluated machine-learning architectures exhibited a performance gradient that was specific to their inherent differences in their ability to describe the degradation processes in sour-service applications of OCTG. Random Forest achieved an average degree of prediction fidelity (accuracy 0.85 ± 0.03; R² = 0.82 ± 0.04), as expected, because tree-based ensembles have been shown to be highly predictive with tabular, non-sequential features, but ineffective with highly path-dependent material degradation processes [8]. This constraint is particularly pertinent to hydrogen-aided fatigue and corrosion-based damage, where sudden degradation does not occur.

Table 3. Comparative performance of machine learning models for failure mode classification and remaining useful life prediction

XGBoost performance was better (accuracy 0.88 ± 0.02; R² = 0.87 ± 0.03), and this result is consistent with the results of other researchers who found that gradient-boosted decision trees outperform traditional ensembles under the condition of nonlinear and multi-interaction relationship of features [9]. Nonetheless, similar to Random Forest, XGBoost still assumes that input samples are independent observations; therefore, it cannot capture temporal aspects, such as crack-growth curves or EIS evolution paths, that control the remaining useful life.

The Physics-Informed LSTM had the greatest predictive power (accuracy 0.93 ± 0.02; R² = 0.92 ± 0.03), which reflects the overall advantage of recurrent neural network sequence modelling together with intrinsic physics-based constraints. This finding confirms the emerging belief that physics-informed architectures are highly effective for improving extrapolation robustness and reducing unphysical material degradation predictions [32, 33]. The LSTM block is trained on long-range temporal dependencies between crack growth histories and the electrochemical evolution of electrochemical evolution data and enforces physics-based penalty terms, which enforce obedience to the mechanistic laws of fracture fatigue kinetics, Arrhenius corrosion acceleration, etc., into a dual constraint learning environment.

Furthermore, the better-than-physical-model performance is also compatible with recent trends in the digital twin literature that discuss the need to adopt a physically grounded AI that ensures reliable functionality in safety-critical systems [31, 34]. Operational decision-making is also critical because its capability of making stable and interpretable predictions is critical in areas where maintenance scheduling and risk mitigation rely on precise remaining useful life (RUL) estimations.

Overall, the performance ranking, Random Forest < XGBoost < Physics-Informed LSTM, is not only experimentally justified in this study but also theoretically consistent with the modern information on degradation modelling, which proves that physics-guided sequential networks provide the most successful basis for OCTG digital-twin integration.

3.5 Digital twin validation and deployment

The verifiability of the physics-informed digital twin was demonstrated by comparing the AI-driven degradation prediction with experimental fatigue, corrosion, and creep data under sour conditions and high temperatures. The additional forecasts of the remaining useful life were not more than approximately 10 percent of the measured values, and the uncertainty bands were able to absorb most of the experimental variations, which presents good model robustness and generalization. The combination of mechanical, electrochemical, and microstructural inputs to make physically consistent predictions and the inclusion of the Paris law and Arrhenius kinetics to avoid non-physical behavior in sparsely populated operating regimes have been achieved [21, 22].

The digital twin translates the predicted rate of degradation into an index of Pipeline Health that is constantly updated, using which real-time maintenance decisions are made. To comply with the dynamic synchronization concepts (as reported by Meza et al. [34]), when the index drops below a certain threshold, automated inspection alerts are produced. Sensitivity analysis, which assessed the modification of the cyclic stress amplitude and partial pressure of H₂S, produced the highest enhancements in the RUL, which is consistent with the findings of sour-service degradation studies [6, 11]. In general, the validated framework is a reliable reproducer of degradation trends and a useful decision-support tool for condition-based maintenance of OCTG systems that can bridge the gap between laboratory characterization and field-scale monitoring of complex service conditions.

The authors gratefully acknowledge the support of the University of Anbar and Mustansiriyah University.