Integrating machine learning, density functional theory, and molecular
A critical review of computer-based research on corrosion finds that combining machine learning, density functional theory, and molecular dynamics could speed up the design of corrosion inhibitors. However, current systems are disconnected, lack enough data, and have not been tested well enough to be trusted for industrial use. The review mainly examines organic inhibitors for mild steel in acidic environments. These environments are found in processes that use hydrochloric and sulfuric acids, such as pickling, acid cleaning, petroleum refining, and chemical processing. The review does not present a new inhibitor. Instead, it examines how the three computational methods can work together and identifies the experimental, computational, and data standards needed to turn computer predictions into candidates supported by laboratory evidence.
Corrosion of mild steel in acid happens through an electrochemical cycle. Iron dissolves at areas called anodic sites, while hydrogen ions gain electrons and form hydrogen gas at cathodic regions. Chloride and other aggressive ions can weaken protective layers on the surface, change the electrical charge at the boundary, and cause damage in specific areas. The boundary between the steel and the solution is constantly changing rather than staying uniform. Traditional methods, such as electrochemical impedance spectroscopy, potentiodynamic polarization, and weight-loss measurements, provide important information about how well an inhibitor works. However, they only reveal molecular behavior indirectly. The review explains that this creates a gap between seeing corrosion protection and understanding how it happens. Researchers often simplify processes such as competing adsorption, changes in protonation, surface changes, and protective film formation, even when a compound appears to work well.
Organic corrosion inhibitors are among the most widely studied ways to protect metals in highly acidic environments. Their performance is often linked to atoms such as nitrogen, oxygen, and sulfur, and to functional groups that help molecules attach to metal surfaces and form protective films. Recent research increasingly focuses on designing molecules based on quantum-chemical measurements and data-driven methods. Materials being studied include amino acids, triazole derivatives, Schiff bases, ionic liquids, plant extracts, polymers, and systems that use nanomaterials. The review points out that these materials may attach to surfaces at several points, work together with other substances, and protect metals through a mix of physical and chemical processes. Because of this complexity, traditional electrochemical measurements alone may not fully explain how some newer inhibitors work.
Density functional theory, or DFT, provides information about the electronic properties of molecules in the proposed workflow. It calculates properties related to how molecules donate or accept electrons and attach to surfaces. These include the energies of the highest occupied and lowest unoccupied molecular orbitals, the energy gap between them, dipole moment, chemical hardness, softness, chemical potential, electronegativity, and electrophilicity. Researchers can combine these measurements with experimentally measured inhibition efficiency to train predictive models. This makes it possible to screen many more molecules than laboratory experiments alone could handle. Many studies link higher highest occupied molecular orbital (HOMO) energies, smaller energy gaps, and related measures of chemical reactivity to better corrosion protection. However, these relationships depend on the dataset. They cannot be assumed to apply in the same way to different groups of inhibitors or different corrosive environments.
The main weakness of DFT is that it usually provides a snapshot of a system rather than showing how it changes over time. Calculations performed in the gas phase or with simplified solvent models may leave out important factors, including how molecules gain or lose protons, how water is arranged, how chloride ions compete for surface sites, and how adsorption changes over time in acidic solutions. Maps of the highest occupied and lowest unoccupied molecular orbitals should therefore be treated as general clues about electronic behavior, not direct predictions of how strongly a compound will attach to a surface under real operating conditions. The review recommends calculations that include individual solvent molecules and the correct protonation states. Researchers should also check experimentally whether the proposed electronic mechanism agrees with electrochemical measurements and observations of the surface. Even with these improvements, DFT alone cannot show whether a protective film will remain stable over a long period or how an inhibitor will behave on an industrial surface that keeps changing.
Molecular dynamics, or MD, adds information about time and the metal–solution boundary. Simulations of metal surfaces, often including the Fe(110) surface, can track how inhibitor molecules move, rotate, attach to the surface, arrange themselves into films, and compete with hydronium or chloride ions for surface sites. Including individual water molecules and ions allows researchers to study interactions involving the solvent, molecules lying nearly flat against the surface, attachment through atoms such as nitrogen or oxygen, and the formation of protective films through cooperation between molecules. These processes are difficult to understand from static calculations alone. MD can also produce numerical measurements, such as adsorption energy, binding distance, and stability at the interface, that can be used in machine-learning models. However, the simulations reviewed in the paper often ran for only hundreds of picoseconds to about one nanosecond, used simplified force fields, and tested only a limited range of temperatures and concentrations. These short simulations can show the early stages of adsorption, but they are not long enough to describe slow surface changes, molecules leaving the surface, or long-term film stability with confidence.
Machine learning provides a way to make predictions across much larger numbers of molecules. Methods such as random forests, support vector regression, artificial neural networks, gradient-boosting models, and newer graph-based methods can connect molecular structures, quantum-chemical measurements, properties at the metal–solution boundary, and measured inhibition efficiency. They can then rank possible inhibitors that would be impractical to test one by one. Tools for measuring feature importance, such as SHAP, can show which properties influence a model's predictions. Clustering methods can group molecules with similar structures. The review also examines a high-throughput process that converts amino-acid structures into computer-readable formats, uses droplet-microarray experiments to map corrosion responses, and combines interpretable models with DFT calculations and electrochemical tests. This repeated process of prediction, testing, and improvement could help researchers use limited laboratory resources more effectively, provided the results remain consistent with physical evidence.
However, the reviewed models often perform well on small, specific datasets but struggle to make reliable predictions for new situations. Training datasets may contain only a few dozen to a few hundred measurements. Reports may also leave out important details, such as the type and concentration of acid, inhibitor dose, exposure time, temperature, steel composition, surface preparation, or the electrochemical method used to measure the target value. Units and reporting formats also differ between studies. Random-forest feature rankings can change when the distribution of molecular properties changes. The performance of support vector models depends on the selected kernel, a mathematical function used by the model. Deep-learning networks are often trained on datasets that are too small and chemically limited to support their complexity. Measures such as mean absolute error, root mean squared error, and R² describe how closely predictions match data or how much prediction error there is. But these measures do not prove that a model follows physical principles, remains reliable when conditions change, or works for molecules outside its training data.
The review identifies the lack of suitable datasets as the most urgent problem. Experimental results on corrosion are scattered across research papers and supplementary files. General chemistry and materials databases contain information about molecules and structures, but they do not provide standardized corrosion measurements. No existing source brings together experimental inhibition efficiency, complete test conditions, DFT descriptors, and MD-derived surface properties in a standardized format specifically designed for corrosion research. The proposed solution is an open benchmark database with consistent units, clearly reported uncertainty, independent test datasets, and data splits that keep related inhibitor families or chemical conditions separate between training and testing. Models should be tested on acids, alloys, and inhibitor classes they have not seen before. Uncertainty should also be tracked through every computational stage instead of being reduced to a single accuracy score.
The case studies discussed in the review show the difference between promising predictions and practical industrial value. A combined machine-learning, DFT, and MD approach used to screen benzimidazole derivatives predicted inhibition efficiencies above 98%, but cross-validation results were only moderate and experimental testing was limited. Other studies reported very high R² values, protection above 85%, or inhibition efficiency of 95%, and used microscopy to confirm protective films. However, these results were limited to particular ionic liquids, concentrations, exposure times, or chemical families. Graph neural network (GNN) and ensemble models reportedly achieved R² values above 0.90 on standardized datasets, but their performance dropped for flexible molecules and fragmented data. These findings come from individual studies. They do not prove that a fully integrated system has already been validated for industrial use.
Another problem is that the software tools used in these studies are often disconnected. Researchers typically transfer information manually between quantum-chemistry software, molecular-dynamics programs, and machine-learning tools written in Python. This makes it difficult to reproduce a complete workflow and limits the ability to combine information across different scales. Computing costs are also important. DFT calculations become expensive for large molecules and extended metal–electrolyte interfaces. MD requires longer simulations and explicit solvent molecules to study adsorption dynamics reliably. The review recommends modular tools with shared interfaces, open workflow templates, and faster substitute models, including machine-learned interatomic potentials, to reduce computing costs. These improvements must maintain consistency in units, physical assumptions, and experimental labels rather than simply speeding up separate calculations that do not work together.
The authors favor a complete discovery system that repeatedly connects simulations and experiments. In this system, simulations guide laboratory tests, validated measurements are added back to the dataset, and models that account for uncertainty choose what to test next. Physics-informed and explainable models could connect predictions to adsorption energy, molecular orientation, competition between ions, and surface coverage. Generative methods could suggest new molecules that meet performance goals and chemical requirements. The review also discusses possible future uses of sensor data in digital twins, which are computer models of real systems, to help adjust inhibitor formulations and doses. However, this is a potential future application, not an industrial capability that has already been demonstrated. The same limitation applies to screening on an industrial scale: the proposed workflow is a plan for future development, not a validated discovery platform.
Overall, the review concludes that no single method can reliably design corrosion inhibitors on its own. DFT provides detailed information about electronic properties but has limited ability to capture changes over time. MD describes behavior at the metal–solution boundary but is restricted by its simulation conditions. Machine learning can make predictions at a large scale, but it depends on the quality of the data, how clearly its results can be explained, and whether predictions are tested independently. In the near term, progress depends less on adding another algorithm and more on making the physical models, experimental methods, and data systems work together. A truly integrated workflow would directly account for solvent effects, protonation, and competition between molecules and ions for surface sites. It would test predictions across several acidic environments, include long-term and industrial conditions, and keep a traceable record from the initial ranking of candidates to experimental confirmation. The review provides a practical plan rather than a finished discovery platform, but it describes a clear path toward designing corrosion inhibitors with less trial and error and stronger support from physical evidence.
Subject of Research: A critical review of how machine learning, density functional theory, and molecular dynamics can be combined to design corrosion inhibitors
Donate to Research
Article Title: Integrating machine learning, density functional theory, and molecular dynamics for smart corrosion inhibitor design: a critical review
Article References: Mohammed, N. J., Yusop, M. R., & Othman, N. K. (2026). Integrating machine learning, density functional theory, and molecular dynamics for smart corrosion inhibitor design: a critical review. Journal of Saudi Chemical Society, 30(3), Article 41. https://doi.org/10.1007/s44442-026-00089-3
Image Credits: Created using artificial intelligence
DOI: 10.1007/s44442-026-00089-3
Keywords: corrosion inhibition, mild steel, machine learning, density functional theory, molecular dynamics, computational chemistry, predictive modeling, acidic corrosion, materials informatics, quantum chemistry, Integrating, machine