Method for predicting efficiency of metal corrosion inhibition based on machine learning model for quaternary ammonium salt compounds

Abstract

FIELD: chemoinformatics and protection of metals from corrosion.SUBSTANCE: method for predicting the efficiency of inhibiting metal corrosion for compounds of the class of quaternary ammonium salts is disclosed, including the steps of forming a data set comprising molecular descriptors of the compounds, their concentration and experimental values of the efficiency of inhibiting metal corrosion, training a machine learning regression model and using it for forecasting, wherein a model based on the CatBoostRegressor gradient boosting algorithm is used as the regression model, wherein before training the model, a feature engineering step is carried out, at which derived features are generated for the inhibitor concentration and the set of molecular descriptors by converting them into a square, cube, logarithm of the form log(1+x) and the inverse of the form 1 / (1+x), and the division of the data set into training and test samples is performed stratified by quantiles of the inhibitor concentration, the set of molecular descriptors for which derived features are generated includes PEOE_VSA4, Chi4n, Inertial Shape Factor, HOMO, LUMO energies, the difference Between the HOMO, LUMO, and TPSA energies, training the CatBoostRegressor model involves a hyperparameter optimization step using a randomized search method with cross-validation.EFFECT: increased accuracy, reliability and generalization ability in predicting the effectiveness of metal corrosion inhibition.1 cl, 5 dwg

Description

[0001] 1. Field of technology to which the invention relates

[0002] The present invention relates to the fields of physical chemistry, materials science, and chemoinformatics, specifically to methods of computer modeling and predicting the physicochemical properties of substances. Specifically, the invention relates to a method for quantitatively assessing and predicting the corrosion inhibition effectiveness of quaternary ammonium salt compounds using machine learning methods.

[0003] 2. State of the Art

[0004] Metal corrosion is a global problem, causing significant economic and environmental damage. One of the most effective methods for combating corrosion is the use of inhibitors. However, the experimental selection and synthesis of new, effective inhibitors is a labor-intensive and expensive process. To accelerate this process, computer modeling methods are actively used, particularly quantitative structure-property relationship (QSPR) models based on machine learning (ML) algorithms.

[0005] The current state of this field is analyzed in detail in a recent review study [Baskin I., Ein-Eli Y. Chemoinformatics for corrosion science: Data-driven modeling of corrosion inhibition by organic molecules / / Molecular Informatics. 2024. Vol. 43. P. e202400082]. The authors of the review highlight several key problems characteristic of most existing approaches:

[0006] The Problem of Small and Homogeneous Datasets: Most published QSPR models are built on small (less than 50 compounds) and chemically narrow datasets. Such models, although they may demonstrate high quality metrics, have a limited range of applicability and are unable to make reliable predictions for structurally diverse compounds.

[0007] Concentration Dependency Problem: The most common target variable, inhibition efficiency (IE%), has a fundamental flaw—it is a function of inhibitor concentration. As Baskin and Ein-Eli note, comparing IE% values ​​measured at different concentrations is invalid, and models that do not include concentration as an input feature are limited in their applicability.

[0008] The need for complex nonlinear models: Metal corrosion inhibition mechanisms are complex and multifactorial, requiring the use of nonlinear machine learning methods capable of capturing complex relationships. This review identifies modern ensemble methods, such as gradient boosting (GBR, XGBoost), as powerful tools for solving such problems.

[0009] The closest analogue (prototype) of the claimed invention can be considered the work [Diah Rahayu Ningtias, Muhamad Akrom. XGBoost Performance in Predicting Corrosion Inhibition Efficiency of Benzimidazole Compounds. Journal of Multiscale Materials Informatics 1(2), 2024, 9-13]. In this work, a method for predicting the inhibitory efficiency of benzimidazole compounds using the XGBoost model is proposed. The model was trained on a dataset of 20 compounds and showed high metrics (R 2 =0.99).

[0010] However, this approach fully reflects the above-mentioned systemic shortcomings of the state of the art.

[0011] Narrow chemical applicability: The model is based on a very small and chemically homogeneous set of 20 benzimidazole derivatives. Its predictive ability for other chemical classes, particularly the widely used quaternary ammonium salts (QAS), is extremely low.

[0012] Ignoring concentration dependence: The prototype model does not use the inhibitor concentration as an input feature, which makes it unable to predict the effectiveness depending on application conditions and limits its practical value.

[0013] Lack of feature engineering: The prototype uses only basic quantum chemical descriptors, without generating derived features to model complex nonlinear effects, which may lead to overfitting on small samples.

[0014] Therefore, there is a pressing technical challenge to create a more universal and efficient prediction method applicable to a broad and practically significant class of inhibitors that would adequately take into account the multifactorial nature of the inhibition process, including the nonlinear dependence on concentration, and would use modern approaches in the field of machine learning for this purpose.

[0015] 3. Disclosure of the invention

[0016] The objective of the present invention is to create a method for predicting corrosion inhibition efficiency that is free from the disadvantages of known solutions.

[0017] The technical result achieved by implementing the invention is an increase in the accuracy, generalizing ability and expansion of the scope of applicability of the predictive model to a practically important and structurally diverse class of inhibitors - quaternary ammonium salts.

[0018] The specified technical result is achieved due to the fact that the method for predicting the efficiency of inhibiting metal corrosion for compounds of the class of quaternary ammonium salts includes the stages of collecting and preparing data, training a machine learning model and using it for prediction, where a regressor based on the CatBoostRegressor gradient boosting algorithm is used as a machine learning model, while before training the model, a feature engineering stage is carried out, at which, for the concentration of the inhibitor and the set of molecular descriptors, derived features are generated by calculating at least their square, cube, logarithm of the form log(l+x) and the inverse value of the form 1 / (1+x), and the division of the original data set into training and test samples is performed stratified by quantiles of the inhibitor concentration.

[0019] 4. Inventive level

[0020] The stated technical result is achieved not by a simple set of known operations, but by their specific, interconnected application, adapted to solving the complex problem of forecasting in the "structure-concentration-property" system. The essence and advantages of the proposed solution, compared to known approaches, are revealed in three points:

[0021] 1. Integration of physicochemical regularities into the model through feature engineering. Unlike standard approaches, where features are generated formally, in the proposed method, the choice of specific mathematical transformations (x 2 , X 3, log(l+x), 1 / (1+х)) is due to the need to approximate known physicochemical dependencies characteristic of metal corrosion inhibition processes: - accounting for adsorption processes. The use of logarithmic and inverse functions allows the model to explicitly take into account the adsorption nature of the process, described by Langmuir-type isotherms, where the efficiency asymptotically tends to saturation with increasing concentration. This enables the CatBoost algorithm to effectively model the dose-response relationship without the need to complicate the structure of the decision trees themselves; - modeling of optimal values. The inclusion of quadratic and cubic functions allows modeling parabolic dependencies characteristic of situations where there is an optimal value of a structural parameter (e.g., the length of an alkyl chain), at which maximum efficiency is achieved.

[0022] Thus, the model initially receives information about the physical nature of the process, which increases its accuracy and interpretability compared to models that operate only with the original descriptors.

[0023] 2. Addressing Estimation Bias by Stratifying by Concentration. Known approaches typically employ standard random sample splitting. However, in the context of this problem, where concentration is the key experimental parameter, this approach introduces a significant technical drawback: the risk of creating unbalanced training and test sets, where the concentration distributions may differ significantly. This leads to a biased model incapable of accurate extrapolation and an unrepresentative assessment of its quality. The proposed method addresses this issue by requiring stratification by concentration quantiles. This step ensures a representative representation of all concentration ranges in both the training and test datasets. This is a prerequisite for constructing a viable model and obtaining an objective, unbiased assessment of its predictive ability.

[0024] 3. Achieving a combination of high accuracy and broad applicability. The proposed combination of steps results in a significant improvement in predictive modeling performance compared to the state of the art. While known solutions demonstrate either high accuracy on narrow, homogeneous sets of chemical compounds (as in the prototype) or lower accuracy on structurally diverse data, the claimed method enables the creation of a model that simultaneously combines high predictive accuracy (R 2 ≈0.83) with a broad applicability range (more than 120 structurally distinct QACs across a wide concentration range). This capability is a direct consequence of the synergistic effect of targeted, concentration-aware feature engineering and a rigorous validation methodology, which together enable the CatBoost algorithm to maximize the efficiency of the provided data.

[0025] 5. Brief description of the drawings

[0026] Figure 1 shows a graph comparing the model-predicted values ​​of inhibitory efficiency with the actual experimental values ​​on the test sample.

[0027] Fig. 2 shows the histogram of the distribution of model prediction errors.

[0028] Figure 3 shows the SHAP summary plot.

[0029] Figure 4 shows a graph of the distribution of errors by inhibitor concentration quantiles.

[0030] Fig. 5 shows the learning curves of the model.

[0031] 6. Implementation of the invention

[0032] The claimed method is implemented by sequentially performing the following steps.

[0033] Stage 1. Dataset formation and characterization. The predictive model is based on a comprehensive and chemically diverse dataset compiled from literature sources and our own experimental studies. The initial dataset includes information on approximately 120 unique quaternary ammonium salts (QAS).

[0034] A key feature of the dataset is its structural diversity, which distinguishes it from the homogeneous samples used in most prior art studies. The dataset includes compounds representing various subclasses of QACs, including:

[0035] - CHAS with long alkyl radicals, for example, dodecyl derivatives, which have pronounced hydrophobic properties;

[0036] - fluorinated QACs containing polyfluorinated alkyl chains (e.g. octafluoropentyl or hexadecafluorononyl fragments), which impart unique steric and electronic properties to the molecules;

[0037] - CHAS containing heterocyclic fragments such as piperazine or imidazoline rings capable of specific coordination on the metal surface;

[0038] - symmetrical dimeric QAS (so-called "twins" or gemini-surfactants) containing two quaternary nitrogen atoms connected by a spacer group. Such compounds are known for their high surface activity;

[0039] - QAC with additional functional groups such as hydroxyl (-OH), ester (-COO-), ether (-O-) and amide (-CONH-), which affect the solubility and adsorption mechanism.

[0040] Another key difference is the inclusion of concentration dependence. For many compounds in the dataset, inhibitory potency measurements are available at several different concentrations, typically ranging from 0.001 mg / mL to 10 mg / mL. This results in a total of six hundred data points (experiments) in the sample, which is sufficient for building a robust model.

[0041] Each data point is described by a set of features including: structural information (IUPAC name of the compound), inhibitor concentration, target variable (percent corrosion inhibition efficiency), calculated molecular descriptors (for each unique QAC cation, an extensive set of descriptors (>50) was calculated using the RDKit chemoinformatics package. This set includes:

[0042] - quantum chemical descriptors - energies of frontier orbitals (HOMO, LUMO), the difference between the energies of HOMO and LUMO (HOMO_LUMO_gap, also called the energy gap);

[0043] - topological indices - Kira-Hall connectivity indices (Chi1n, Chi4n);

[0044] - physicochemical descriptors - molecular weight (MolWt), topological polar surface area (TPSA), hydrophobicity (MolLogP);

[0045] - shape descriptors - asphericity, eccentricity, sphericity index, inertial shape factor;

[0046] - structural prints (MorganFingerprint).

[0047] Stage 2. Advanced Feature Engineering. At this stage, new derived features are generated for the original feature set to capture complex nonlinear relationships. For inhibitor concentration and key molecular descriptors (PEOE_VSA4, Chi4n, HOMO_LUMO_gap, etc.), their square, cube, logarithm (log(l+x)), and reciprocal (1 / (1+x)) are calculated.

[0048] Stage 3. Splitting and preparing training data. The dataset with the original and generated features is split into training (80%) and testing (20%) sets, stratified by inhibitor concentration quantiles. Missing values ​​in numerical features are filled with the mean value.

[0049] Stage 4. Building and optimizing the predictive model. CatBoostRegressor is used as the algorithm. Hyperparameters are selected using the Randomized Search method, a stochastic optimization algorithm that randomly selects combinations of hyperparameters from given distributions using 5-fold cross-validation.

[0050] Stage 5. Training the final model and evaluating its performance. The performance of the trained model is evaluated on the holdout test set. The main quality metrics are:

[0051] - coefficient of determination (R 2 ) = 0.8231;

[0052] - Root Mean Square Error (RMSE) = 8.3894%;

[0053] - mean absolute error (MAE) = 5.3499%.

[0054] Additionally, to confirm the performance, a 5-fold stratified cross-validation was conducted, which showed the following average results:

[0055] - R 2 = 0.8274±0.0533;

[0056] - RMSE = 8.7827±1.2025%;

[0057] -MAE = 5.5711±0.8204%.

[0058] These indicators indicate high predictive ability and stability of the model.

[0059] Stage 6. Analysis and interpretation of results. A thorough visual and statistical analysis of the model's performance is performed.

[0060] The "Actual vs. Predicted" graph (Fig. 1) demonstrates a high correlation between the actual and predicted values. The points are tightly clustered along the diagonal line of perfect fit (y = x), confirming the high R 2 .

[0061] Error histogram (Fig. 2): The residuals are distributed nearly normally, centered at zero. The vast majority of errors are in the range of -10% to +10%, which is characteristic of a well-tuned model.

[0062] Feature importance analysis (Fig. 3): SHAP analysis clearly shows that the most significant factor is the inhibitor concentration and its nonlinear derivatives. This experimentally proves that addressing the concentration dependence problem through advanced feature engineering is a key factor in the success of the model.

[0063] Error analysis by concentration quantiles (Figure 4): The graph confirms the model's performance across various concentration ranges. The median error for all five concentration quantiles is close to zero. This proves that the model performs equally well at both low and high concentrations.

[0064] Learning curves (Figure 5): The graph shows that the error on the validation set decreases rapidly and reaches a plateau. The small gap between the training and validation curves indicates good generalization ability and the absence of overfitting.

[0065] Step 7. Model Application. The trained model is saved and used for virtual screening of new candidate compounds of the QAC class.

Claims

A method for predicting the efficiency of inhibiting metal corrosion for compounds of the class of quaternary ammonium salts, including the steps of forming a data set containing molecular descriptors of the compounds, their concentration and experimental values ​​of the efficiency of inhibiting metal corrosion, training a machine learning regression model and using it for forecasting, characterized in that a model based on the CatBoostRegressor gradient boosting algorithm is used as a regression model, while before training the model, a feature engineering step is carried out, at which derived features are generated for the inhibitor concentration and the set of molecular descriptors by converting them into a square, cube, logarithm of the form log(1+x) and the inverse of the form 1 / (1+x), and the division of the data set into training and test samples is performed stratified by quantiles of the inhibitor concentration, the set of molecular descriptors,for which derived features are generated, includes PEOE_VSA4, Chi4n, Inertial Shape Factor, HOMO, LUMO energies, the difference between HOMO, LUMO and TPSA energies, training the CatBoostRegressor model includes a hyperparameter optimization step using a randomized search method using cross-validation.

Images