A Reproducible Statistical and Explainable Machine Learning Workflow for Small Corrosion Datasets: Methodological Demonstration Using Measured Weight-Loss and Synthetic Corrosion-Rate Data

1. INTRODUCTION

Corrosion is a serious industrial problem and is responsible for annual economic losses of about US$2.5 trillion (or 3–4% of the world’s GDP) and endangers critical infrastructure like pipelines, bridges and offshore platforms [1]. This general electrochemical process may promote deterioration of the materials and may lead to environmental contamination via the evolution of metallic species and corrosion products, making it even more important to adopt effective and sustainable protection measures [2]. Environmental factors like humidity, salinity and acidity create acidic conditions that support metal dissolution at the anodes of a corrosion system and the necessary cathodic reactions, thus strongly influencing the corrosion behaviour [3]. Mainstream methods used mainly included conventional inhibitors (chromates and organic amines), but their toxicity and regulatory restrictions have shifted the focus to looking for more environmentally friendly alternatives [4]. The use of nanostructured materials has been widely introduced in the literature as possible methods for preventing corrosion [5]. The corrosion resistance of copper (Cu), nickel (Ni) and chromium (Cr) coated materials have been reported as improved in some experimental media with inhibition efficiency up to 98.5% in aggressive media [6]. They have shown potential to have a beneficial effect on the surface stability and corrosion resistance under certain experimental conditions, but the environmental sustainability of these materials depends on the composition, synthesis route, toxicity, service-life behaviour and end-of-life management.

Despite these developments, evaluation methods still face limitations. Conventional laboratory techniques, such as the weight-loss method and potentiodynamic polarisation, can provide valuable information on corrosion behaviour but are time-consuming and cannot fully simulate the heterogeneous environmental, mechanical, and service conditions found in industry over long service periods [7]. This can hinder the efficiency, scalability and repeatability of traditional corrosion-performance testing. Furthermore, numerous studies have investigated the influence of single experimental parameters but have not properly accounted for the combined effects of exposure duration, coating chemistry, and environmental variation, leading to a more partial understanding of corrosion behaviour [8]. One remaining methodological issue is the lack of openly reproducible, data-driven frameworks that combine statistical inference, machine-learning comparison, and explainability for corrosion datasets. It is therefore not a discovery of new corrosion mechanisms, but rather a framework for analytical and reproducibility purposes, which is the contribution of the present study.

While some researchers have used ML to predict corrosion in alloys before Galvão et al. [9], few have used other publicly available databases, such as Kaggle Corrosion Rate, to compare coating performance under similar conditions. In addition, electrochemical impedance spectroscopy (EIS) remains the most common method to assess corrosion; explainable artificial intelligence techniques like SHAP have been studied little to interpret data-driven corrosion models. To fill this methodological void, this paper presents a reproducible pipeline implemented in Python that combines statistical analysis (ANOVA and Tukey HSD), predictive models of machine learning (Random Forest and XGBoost), and techniques from the field of explainable artificial intelligence (XAI) to illustrate a clear framework for processing corrosion-related data [10]. Beyond predictive assessment, the new research paradigm should consider multifunctional smart nanostructured coatings that combine passive barrier protection, stimulus-responsive inhibitor release, and real-time corrosion sensing in a single architecture. A nanostructured matrix in this system may block the flow of water, oxygen, and aggressive ions, while functionalised nanoparticles or nanocontainers may store corrosion inhibitors to release if the coating is damaged or if local pH and/or ionic conditions change. Corrosion-responsive sensing components can sense early electrochemical and/or chemical changes that indicate substrate degradation. Integrating these sensing outputs with IoT-enabled monitoring and machine-learning models could facilitate condition-based maintenance and predictive corrosion management. Therefore, future development of such multifunctional systems should optimize barrier properties, inhibitor release rate, sensing accuracy, coating adhesion, environmental safety, manufacturability, and barrier durability simultaneously.

This research introduces and validates an interpretable and reproducible computational process to statistically infer, predict, and explain small corrosion datasets using an artificial intelligence approach. Only quantities directly supported by the measured source, cumulative bare/coated mass loss and protection efficiency vs exposure time, are used for the measured source. Since no density of specimens nor area of exposed specimens are provided, the corrosion rate is not measured and reported directly in mm/year; no identification of the coated specimen as Cu, Ni, or Cr is provided, so no claim for experimental validation of coating in relation to copper, nickel, or chromium. The synthetic data set is not intended to predict the corrosion performance of Bare, Cu, Ni, or Cr, but is presented as a controlled computational environment to demonstrate the workflow implementation across coating categories. RQ1 analyses the statistical difference between the pre-defined synthetic coating groups; RQ2 compares the predictive models by performing cross validation and tuned held out tests; RQ3 examines the relative contribution of the coating category and exposure days for the predictive model by using one-hot-compatible feature importance and aggregated SHAP analysis; and RQ4 explores how the workflow can be extended to experimentally proven corrosion datasets.

This research is not concerned with materials’ performance; it is a methodological contribution. The study first presents a reproducible analytical workflow that integrates statistical inference, machine-learning comparison, uncertainty estimation, and explainable artificial intelligence for small corrosion datasets. Second, the study is a strong example of how synthetic datasets can be transparently built and analysed, even when little publicly available corrosion data exists. Third, the workflow highlights the need for the right encoding approach, validation methods, and explainability when processing heterogeneous corrosion datasets. Last but not least, the study outlines needs for future application by applying validated experimental corrosion data. The results show that a priori synthetic data yield a lower corrosion rate for Cr and Ni than for bare steel, and that the simulated mean corrosion rate for Cr is the lowest among the coating categories analysed.

The paper is organized in such a way that in Section 2, there is a review of literature on the topic of inhibitors and ML applications, Section 3 explains the methodology, including data preprocessing, EDA, statistics, and ML workflows, Section 4 presents the results of the statistical tests, models performances, and visualizations, Section 5 provides interpretations, alignments to mechanistic understanding, and limitations, in Section 6, the paper outlines challenges and future directions, and provides implications of scalable frameworks.

2. LITERATURE REVIEW

2.1. Conventional Corrosion Inhibitors

Traditional corrosion inhibitors have been the most important means of protecting metals from corrosion, and they include organic, inorganic, and mixed chemical compositions and modes of action [11]. Organic inhibitors may be amines, carboxylates, or heterocyclic molecules containing N, O, and/or S atoms and act predominantly by forming protective films on the metal surface, which retard both anodic and cathodic reactions [12]. Under specific acidic conditions, these inhibitors can undergo physical or chemical adsorption on metallic surfaces, with their adsorption behaviour often following a model (such as the Langmuir or Temkin isotherm). Some conventional organic inhibitors may have issues with biodegradability, persistence, and potential toxicity; however, these can restrict their large-scale use.

Anodic corrosion inhibitors such as chromates, nitrites, and phosphates can form or stabilize surface films that protect against corrosion [13]. Chromate-based inhibitors have historically offered good corrosion protection but are now increasingly limited by regulations such as REACH because of the toxic and carcinogenic nature of hexavalent chromium compounds [14]. The mixed inhibitors are composed of both organic and inorganic components that act synergistically on both half-cell reactions to improve performance in a medium of neutral or alkaline pH [15]. Although these conventional inhibitor systems work well, they often require relatively high concentrations, can persist in the environment, and can perform inconsistently under varying pH, temperature, and exposure conditions, prompting the need to investigate alternative systems using nanostructures.

2.2. Nanostructured Inhibitors and Coatings

In recent years, the emergence of new nanostructured materials has increased the possibilities for corrosion protection by offering a high specific surface area and tuneable interfacial properties, which can be used for barrier reinforcement, surface modification, and active delivery of corrosion inhibitors [16]. Nanostructured materials do not have to replace the currently used organic or inorganic inhibitors but can be used in combination in synergic protection systems. Functionalised nanoparticles and nanocontainers could also serve as reservoirs or transporters for organic inhibitor molecules, releasing the inhibitor molecules in a controlled or stimulus responsive manner upon changes in pH, ionic concentration and/or integrity of the coating induced by local corrosion. The inorganic inhibitors can also cause substrate passivation and the nanostructured fillers can also lower coating permeability and limit transport of corrosive species. This hybrid protection can then be a passive barrier protection plus active corrosion suppression once local coating damage occurs. However, to achieve true synergy, optimization of the concentration of nanoparticles, loading of inhibitors, release kinetics, chemical compatibility, coating adhesion, mechanical properties, environmental safety, and manufacturing reproducibility are required [17]. Some materials, like zinc oxide (ZnO), titanium dioxide (TiO₂), and composites based on graphene, have been found to possess desirable barrier properties and, if designed properly, some nanostructured systems may also exhibit barrier properties that are active and/or self-healing. The efficiencies of corrosion inhibition of the green-synthesised ZnO nanoparticles have been reported up to 98.5% under certain experimental conditions by forming tight protective films that limit chloride ion transport. Likewise, the TiO₂ nanotubes can photocatalytically decompose corrosive substances and, when applied to the surfaces of steel, form passive oxide films.

Multi-functional coating materials, such as graphene and derivatives of graphene, graphene oxide (GO) and reduced graphene oxide (rGO), have been investigated owing to the fact that their two-dimensional structures can create tortuous diffusion pathways, which will limit the diffusion of gaseous and ionic corrosive species [18]. Metallic nanostructured coating such as Cr and Ni alloy can be deposited by electrodeposition or physical vapour deposition, and exhibit a pitting potential of 300-500 mV higher than the saturated calomel electrode. As an example, Ni-Cr nanocomposites on 316L stainless steel can increase the corrosion resistance of 3.5% NaCl, as evidenced by electrochemical impedance spectroscopy (EIS) which indicated that |Z| > 106 -1 0 cm² at low frequency. While nanostructured coatings might decrease the need for some of the conventional and hazardous inhibitors, their sustainability cannot be taken for granted, as the production of nanoparticles, work-site exposure, release to the environment, persistence, leaching out of metal ions, and disposal at the end of their life could create new environmental and toxicological challenges. The small size and high surface reactivity of metallic and metal-oxide nanomaterials are important for the biological interactions, environmental mobility and ecotoxicity [19]. Nanomaterial strategies based on green and bio-inspired approaches can help lower some of the environmental impacts by using plant-based reducing agents and stabilisers, biodegradable polymers, natural inhibitor molecules and synthesis methods that avoid hazardous chemicals. These methods can not only diminish the use of harmful chemicals, but can also maintain the surface activity and barrier properties critical for corrosion protection. However, the use of a nanostructured inhibitor as ‘green’ should not be based only on the efficiency of corrosion-inhibition, but also on the results of the toxicity, leaching, biodegradation, life-cycle and end-of-life assessments [20]. Comparative characteristics of carbon-based, metal-oxide and ferrite nanostructures.

2.3. Comparative Characteristics of Carbon-Based, Metal-Oxide, and Ferrite Nanostructures

Carbon based nanostructures, such as graphene, carbon nanotubes (CNTs), and carbon dots have large specific surface areas, tuneable surface chemistry and the potential to produce tortuous diffusion pathways that prevent the penetration of water, oxygen and aggressive ions through protective coatings [21, 22]. The reinforcement of barrier layers with graphene-based materials is of special interest while functionalised CNTs could help with mechanical reinforcement as well as active incorporation of inhibitor species that form part of the material [23]. But agglomeration, bad dispersion, incompatibility at the interface and complexity of processing can be problematic with carbon-based nanostructures, and in using conductive carbon materials, proper design of the interface with the metallic substrate can be a challenge. Their persistence in the environment and potential exposure to the workplace must also be taken into account [24, 25].

The metal-oxide nanoparticles (ZnO, TiO₂) have various advantages: chemical stability, easy synthesis, high interfacial activity, and the ability to form additional functional properties and passive film [26]. An important feature of nanoscale ferrite material is its intrinsic magnetic responsiveness that is very appealing for architectures of multifunctional coatings where corrosion protection is achieved in combination with magnetic properties [12, 27]. On the other hand, metal-oxide and ferrite nanoparticles can also cluster, release ions, leach and have particle-size dependent particle performance with often strong dependency on coatings matrix compatibility as well as on surface functionalisation. Thus, the choice of materials between carbon and metal-oxide/ferrite nanostructures should depend on the metallic substrate, corrosive conditions, coating chemistry, desired functionality, environmental impact, scalability and cost, and should not be based on the assumption that one class of nanomaterials is superior to the other.

2.4. Data-Driven and Machine Learning Approaches in Corrosion Prediction

To predict the corrosion rate of steels and alloys, based on the combination of environmental, microstructural and electrochemical parameters, data-driven techniques, particularly Machine Learning (ML) have been developed, with better prediction accuracy. Random Forest (RF) and gradient boosting models like XGBoost have also performed well with R2 close to 100% for predicting atmospheric corrosion as they can model the nonlinear relationships [28, 29]. In dynamic environments, such as concrete-embedded rebar, the support vector regression and neural networks have been used to determine the features that most affect the prediction of the alloy corrosion, such as chloride content and pH. Recent studies with 3D-printed micro-lattices demonstrate the possibility of using ML to correlate surface morphology and corrosion, and the XGBoost has been able to determine the important hyperparameters that enhance the overall generalisation [30]. Explainable AI, like SHAP, provides explanations of features of predictive performance, like the nonlinear impact of exposure time, in line with Tafel kinetics. However, many applications have proprietary data sets, limiting the reproducibility [31].

2.5. Research Gaps

There is still a methodological void in open-source corrosion research for the integration of the use of statistical inference, machine-learning prediction, and explainable artificial intelligence (XAI) in a reproducible analytical framework. The studies conducted so far are mostly experimental electrochemical or laboratory data; there has been relatively little work that uses publicly available datasets that can be used to demonstrate the statistical or machine-learning workflows in a reproducible manner. Few reproducible studies combine hypothesis testing, predictive model comparison, and explainability into a cohesive corrosion-analysis workflow, especially in the context of interpreting categories of nominal materials or coatings, without artificially imposing an ordinal ordering. There are thus opportunities to put into place reproducible Python-based frameworks that encompass statistical analysis, machine learning, and explainability with a dedicated examination of differences between synthetic and empirical corrosion datasets and to extend the work to multivariate industrial datasets. Solving this methodological limitation would help to enable more transparent and reproducible evaluation of the corrosion related datasets, and could serve as a template that could then be tested with experimental data.

3. METHODOLOGY

3.1. Dataset Description

The source dataset includes 13 measured exposure-time observations (0, 5, 10, …, 60 days) for a bare steel specimen and one generically coated specimen. Based on these measurements, the study calculated cumulative mass loss (g), cumulative mass loss (%), and protection efficiency (%). The source did not provide the specimens’ density or exposure area, so the standard conversion (weight loss to corrosion rate) cannot be done without unsupported assumptions. It is also not specific to Cu, Ni, or Cr, and therefore can’t be validated using metadata of the measured Cu/Ni/Cr generic coated specimen. A synthetic n = 100 dataset was generated separately with NumPy (random_state = 42) for methodological demonstration. Synthetic corrosion-rate values were generated from predefined normal distributions: Bare N (0.50, 0.05), Cu N (0.18, 0.03), Ni N (0.13, 0.02), and Cr N (0.10, 0.02) mm/year, while exposure days were sampled from {0, 5, 10, 15, 20, 30, 45, 60}. To ensure separation of coating categories with a generally kept within-group variability, the mean and standard deviation of the corrosion rates used for the synthesis were chosen as pre-defined simulation parameters. Both selected means and standard deviations were chosen to be hypothetical differences between the unprotected and protected condition; the standard deviations were chosen to avoid perfectly deterministic difference between the groups. These values are not determined by experimental Cu, Ni or Cr coating studies, and are used as a computation guess for the coating properties for workflow evaluation. They are simulation assumptions and not measurements of the coating performance. Inferential statistics, machine learning accuracy, feature importance and SHAP values for the synthetic dataset then measure relationships in this synthetic analytical dataset and are not estimates of the effectiveness of the coatings in the real world. The methodological process is shown in Fig. (1).

Fig. (1). Proposed methodology.

CONCLUSION

This study demonstrates a revised, reproducible workflow that distinguishes measured weight-loss evidence from synthetic corrosion-rate modelling. The measured source supports cumulative mass-loss analysis and a Day-60 protection efficiency of 88.89% for a generic coated specimen, but direct measured corrosion rate in mm/year cannot be calculated without specimen density and exposed area, and Cu/Ni/Cr-specific experimental validation cannot be established without coating metadata. In the synthetic dataset (n = 100), ANOVA confirmed strong predefined between-group separation (F = 744.7167, p = 2.524017 × 10⁻⁶⁶, η² = 0.9588). With nominal coating type represented by OneHotEncoder, Linear Regression achieved the highest mean five-fold cross-validation R² (0.9517), while tuned XGBoost achieved the highest held-out test R² (0.9175; RMSE = 0.0384; MAE = 0.0253). However, these performance estimates should be interpreted within the context of the limited synthetic sample size (n = 100), where model ranking, feature importance, and predictive stability may vary when evaluated using larger experimentally validated datasets. Random Forest importance was led by Coating_Bare (0.955253), and aggregated SHAP importance was 0.980856 for coating category versus 0.019144 for exposure days. These values reproduce the final implementation outputs and describe only the constructed analytical dataset. Future validation requires measured specimen area and density, explicit Cu/Ni/Cr coating identities, and matched experimental corrosion-rate observations.

LIST OF ABBREVIATIONS

AUTHOR’S CONTRIBUTION

S.S. has contributed to the study conceptualization, methodology, data analysis, interpretation of results, and manuscript writing.

ETHICAL APPROVAL & INFORMED CONSENT

This study did not involve human participants or animal subjects; hence, ethical approval was not required.

AVAILABILITY OF DATA AND MATERIALS

The datasets generated or analysed during this study are available from the corresponding author on reasonable request. The base data are accessible via the public Kaggle repository (“Corrosion Rate Dataset”). The measured source data and the code required to reproduce the synthetic dataset and analytical workflow are also available from the corresponding author on reasonable request. The synthetic corrosion-rate dataset used for machine-learning demonstration was generated computationally from the distributional assumptions reported in Section 3.1 and was not interpreted as independently measured Cu, Ni, or Cr corrosion data.

FUNDING

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

CONFLICT OF INTEREST

The author declares that there are no conflicts of interest regarding the publication of this paper.

ACKNOWLEDGEMENTS

The author acknowledges Kaggle for providing the open-source corrosion dataset and express gratitude to the open-source Python community for developing the computational tools (NumPy, scikit-learn, SHAP, and Matplotlib) that enabled this research.

DECLARATION OF AI

During the preparation of this manuscript, the author used ChatGPT for language editing and refinement purposes. Following the use of this tool, the author carefully reviewed and revised the content where necessary and accept full responsibility for the final published version of the article.