Artificial Intelligence, Machine Learning, And Big Data For Corrosion Control – Quo Vadis?

Authors: D.A. Winkler, A.E. Hughes, C. Özkan, J.M.C. Mol, T. Würger, C. Feiler, S. Lamaka.

This paper was presented at the Corrosion & Prevention 2023

ABSTRACT

The current era has seen amazing, paradigm shifting scientific developments. We now know that the size of small molecule and materials spaces is for all practical purposes, infinite. In many fields, this recognition has seen a rapid increase in automation and robotics, allowing synthesis of new molecules and materials and measurement of properties much faster than before. Unfortunately, corrosion science has not yet seen the benefit of these advances, but undoubtedly will. Finding ‘islands of chemical utility’ in vast chemical spaces requires AI and data driven machine learning. Machine learning algorithms are universal approximators, applicable to diverse applications, such as medicine, finance, manufacturing, social media etc. Recent developments in quantum machine learning methods, generative methods to design new molecules and materials with improved properties, massive ‘make on demand’ chemical libraries such as the 30+ billion ZINC-22 library, autonomous chemical discovery systems with no human in the loop, and other AI methods (evolutionary algorithms) are accelerating discovery of novel, useful molecules and materials. This paper summarizes technical developments in the corrosion inhibition field. We describe how data-driven machine learning methods generate models linking molecular properties to corrosion inhibition can predict performance of materials not yet synthesized or tested. We provide a perspective on the benefits of additive manufacturing, high throughput corrosion inhibitor testing, other AI methods such as evolutionary algorithms, and autonomous corrosion inhibitor design systems. These exciting technologies represent a paradigm shift in rapid discovery of novel, effective corrosion inhibitors for a range of metals and alloys, in diverse environments.

Keywords: Organic corrosion inhibitor; machine learning, molecular design, chromate replacement, chemical space

INTRODUCTION

The extremely impressive developments in a diverse range of technology such as computational power, machine learning algorithms, high throughput robotic synthesis and characterization of molecules and materials and make[1]on-demand chemical libraries is driving a paradigm shift in innovation in most areas of science, technology, and medicine [1, 2]. These technologies are very synergistic and generate outcomes that are potentially larger than the sum of the parts. The relentless increase in computational power is allowing the development of larger and more complex machine learning methods, high throughput synthesis and characterization is generating very large data sets for training machine learning models, and ready availability of immense chemical libraries is allowing more of chemistry space to be accessed quickly, cheaply, and without the need, necessarily, of in-house chemical synthesis facilities. Wider recognition of the essentially infinite size of chemistry space (1060-10100) suggests that there are much better chemical solutions to problems available than have been found by our exploration of an infinitesimally small fraction of these spaces over the last 100 years. However, navigating chemical space to find these ‘island of utility’ is daunting. Although we can now synthesize and test molecules orders of magnitude faster than before, experiments alone cannot explore significant parts of chemical space now or in the future. Computational methods are required to leverage the experiments into much larger regions of chemical space.

Despite corrosion control costing more than US$1Tn per annum, the corrosion protection industry has been a very late adopter of the technologies discussed above. There are relatively few high throughput testing rigs for assessing the performance of candidate inhibitors, so the data sets available to train machine learning models are very small. This means that even when good models are developed from these data, their domains of applicability (the regions of chemical space in which they make accurate predictions) is limited, hampering their use. There are also relatively few expert uses of machine learning methods in corrosion science, relative to other areas of science and technology. A significant number of models using machine learning to model corrosion inhibitors use very small data sets and the conclusion drawn from the calculations are limited at best and inaccurate at worst. The potential for machine learning and other AI methods to provide a large step change in discovery of effective and safe organic replacements for inorganic corrosion inhibitors, increasingly being restricted or banned because of toxicity, is not being realized. These methods are extremely fast and cost-effective for discovering or designing new corrosion protection agents.

WHAT IS HAMPERING WIDER USE OF MACHINE LEARNING DESIGN METHODS IN CORROSION?

Lack of data is the largest limitation to the wider application of machine learning to discovery of new inhibitors. Interestingly, the same technology is equally applicable to the discovery of modulators of metal dissolution that are invaluable components of next generation battery technologies. The generation of molecular features encoding relevant properties of molecules, optimal selection of these features from a pool of possibilities, generating quantitative structure-property relationship (QSPR) models using machine learning, validating the models to determine their predictive power and domain of applicability are all essentially solved problems in other areas of science that can be applied directly to corrosion science problems.

The problem of generating larger corrosion inhibition data sets has been tackled by several research groups, including our own. Multi-electrode electrochemical methods have been reported by Muster et al. and Garcia et al. [3, 4], with capacities of 30 electrochemical experiments per hour. Recently, Breedon et al. reported and patented an electrochemical robot capable of measuring the corrosion inhibition properties of 80 compounds each session [5]. Optical methods exploiting the brightness of corroded regions of metals have been successfully exploited by White et al. [6] and Shi et al. [7] Lamaka and colleagues developed a medium throughput corrosion inhibitor assessment method for Mg alloys that exploits hydrogen evolution and can measure ~150 compounds per study [8]. These efforts show that higher throughput assessment of corrosion inhibition properties of organic materials is possible, but the scale must be increased by an order of magnitude or more to begin to address the data paucity issues preventing full exploitation of machine learning modelling methods. Importantly, the issue of accelerated corrosion methods predicting the performance of inhibitors under ‘real world’ operating conditions is always present and must be addressed in the design of the high throughput ‘surrogate’ corrosion testing processes.

EXAMPLES OF USE OF MACHINE LEARNING TO MODEL ORGANIC CORROSION CONTROL AGENTS

We summarize selected recent research, largely from our research groups, on the application of machine learning to aluminium and magnesium alloy corrosion inhibitors that use larger data sets [9-14]. Unfortunately, many earlier ML modelling studies of organic corrosion inhibitors were based on data sets of less than 10 compounds, crippling the ability of the ML algorithms to generate robust and broadly predictive models. We have also summarized selected, recent, relevant research on the same alloys by other researchers for balance. For additional information, a recent review has summarized relevant research on applying ML methods to organic corrosion inhibitor data.[15]

CORROSION INHIBITORS FOR ALUMINIUM AND ALLOYS

Galvao et al. reported the results of categorical modelling of inhibition by organic agents (inhibitor/non-inhibitor) by a range of ML methods – k-nearest neighbours, decision trees, decision trees with boosting, decision trees with defined error costs, bagging, random forests, classification rules, artificial neural networks, and support vector machine.[16] Random forest models had balanced prediction accuracies of 82-85% (50% is random guess – i.e., no model), and decision trees (with and without boosting) had balanced prediction accuracies of ~75%. Balanced accuracies account for the training set having more examples in one class than the other. They also identified two new molecular features for modelling corrosion inhibition by organic molecules, : dimerization enthalpies and Gibbs energies from quantum chemical (Gaussian) calculations.

Dibetsoe and colleagues used quantum chemical methods to predict the corrosion inhibition of seven macrocyclic compounds against 100% aluminium in 1M HCl measured by weight loss.[17] The dataset consisted of four phthalocyanines, 1,4,8,11,15,18,22,25-octabutoxy-29H,31H-phthalocyanine, 2,3,9,10,16,17,23,24-octakis(octyloxy)-29H,31H-phthalocyanine, 2,9,16,23-tetra-tert-butyl-29H,31H-phthalocyanine and 29H,31H[1]phthalocyanine and three naphthalocyanines, 5,9,14,18,23,27,32,36-octabutoxy-2,3-naphthalocyanine, 2,11,20,29-tetra-tert-butyl-2,3-naphthalocyanine and 2,3-naphthalocyanine. They developed two linear regression models with three parameters each that exhibited r2 values of 0.94. However, the graphs of predicted versus experimental inhibition contained only four of the seven compounds and, given that 3 parameters were fitted to each model, suggests these models were also overfitted.

High throughput corrosion inhibition data sets for libraries of small organic molecules and AA2024-T3 and AA7075-T6 aerospace alloys were reported by Harvey et al. and White et al. [6, 18] The so-called Harvey data set contained 28 small organic molecules whose corrosion inhibition was assessed by mass loss over 28 days immersion in saline. The larger library of 100 organic agents, reported by Harvey et al., was assessed for corrosion inhibition using a novel optical method. A multiwell plate, where each well contained a corrodent, inhibitor and control, was applied to a sheet of metal under study. The average brightness of the pixels in each corrosion circle correlated directly with the extent of corrosion. After scaling, the brightness values were ranked from 0 (darkest, least amount of corrosion) to 10 (brightest, most amount of corrosion), rounded to the nearest integer. This inhibition of corrosion by this library was also measured at several initial pH values, 3, 4, 5, 6, 8, and 10.

In our ML studies, the molecules in the libraries were converted into mathematical features used to train machine learning models of structure-inhibition relationships. Despite several reports claiming that features derived from quantum chemical calculations (e.g., HOMO, LUMO, and band gap energies etc.) being important for corrosion inhibition of aluminium alloys, using with much larger and more chemically diverse data sets we found no such correlation [12]. Relevant, context-aware subsets of features were selected using a sparse feature selection method, MLREM, which also generated parsimonious multiple linear regression models. The data were also used to train nonlinear machine learning models using Bayesian regularized feed forward shallow neural networks. The neural networks consisted of three layers, an input layer receiving the features, a hidden layer containing a small number of neurons with nonlinear sigmoidal transfer functions, and an output layer generating the predicted inhibition values. Input data were auto scaled and the networks trained to the maximum of the Bayesian evidence for the model using a Levenberg Marquart algorithm. Models were validated by use of a separate test set not used to generate the model.

For the Harvey data set, inhibition results over 28 days immersion in saline for the two alloys AA2024-T3 and AA7075-T6 were strongly correlated (r2=0.84) [11]. Models trained on both inhibition data sets employed 173 in[1]house, and 194 Dragon molecular descriptors, reduced to the 8 most relevant features by sparse feature selection. The linear MLREM and machine learning models were highly statistically significant.

For the AA7076-T6 alloy, the training set was predicted with r2=0.86, standard error=43% and the test set was predicted with r2=0.91and standard error=36%. The nonlinear model was slightly better, predicting the training set with r2=0.77, standard error=24% and test set with r2=0.88, standard error=32%. Inhibition of the AA2024-T3 alloy was more difficult to model, and the training set was predicted with r2=0.83, standard error=46% and test with r2=0.80 and standard error=49% for the MLREM linear model. The performance of the nonlinear neural network model was similar predicting the training set with r2=0.81 and standard error=28% and the test set with r2=0.74and standard error=45%. The ability of the models to predict experimental corrosion inhibition results is summarized in Figure 1.

Machine learning modelling of the larger and more chemically diverse White data set assessed at a concentration of10−3 M in 0.1 M NaCl solutions also yielded statistically significant models with useful predictive power [12]. The inhibition of corrosion by the test compounds was assessed by image processing after a 24-h period. Corrosion extent was accurately estimated from the brightness of the images of the circular regions on the alloy exposed to inhibitor solutions in an 8 × 11 grid of 6 mm diameter grid of cells. The scores ranged from 0 (no corrosion) to 10 (maximum corrosion with no inhibitor). For this much shorter contact time assay, the correlation between the inhibition of the AA2024 and AA7075 alloy by the chemical library was substantially low (r2 of 0.27-0.35). An initial pool of 1515 DRAGON and in-house descriptors was used, reduced to 17 most relevant to inhibition by sparse feature selection. The most robust MLREM model for AA7076 could predict the training set with r2=0.77, standard error=1.2 and the test set with r2=0.79 and standard error=1.1. The neural network models using 2 or 3 neurons in the hidden layer generated models with very similar accuracies, predicting the training set with r2=0.71 and standard error=1.0 and the test set with r2=0.82 and standard error=0.9. For the AA7076 alloy, the model with the best predictive accuracies used 29 descriptors and could predict the training set with r2=0.85 and standard error=1.0 and the test set with r2=0.75 and standard error=1.2. The quality of predictions for both alloys by the most robust machine learning models is illustrated in Figure 2.

Atlam et al. recently reported a quantum chemical analysis of corrosion inhibition of pure aluminium by six amino acids by weight loss and polarization measurements.[19] Using density functional theory (DFT), they calculated a range of molecular properties, the energy of highest occupied molecular orbital (EHOMO), energy of lowest unoccupied molecular orbital (ELUMO), the energy (band) gap (ΔE), global hardness (η), global softness (S), ionization energy (I), electron affinity (A), electronegativity (χ), chemical potential (μ), molecular volume (MV), total energy change (ΔET) and fraction of electron transferred (ΔN); plus local reactivity indices such as Fukui function and dual descriptors. The did not find a linear model relating these molecular features to corrosion inhibition, but generated apparently strong nonlinear models of the following form: –

Where A-F are fitted constants to the model and Cinh is the concentration of the inhibitors. Not surprisingly, the model predicted the experimental corrosion inhibitions extremely well because the model was overfitted (same number of fitted parameters as training examples. The results are therefore not particularly useful for identifying new molecules or amino acids as potential inhibitors. These results are also at odds with those from our study using a much larger and chemically diverse dataset that found no correlation between MO calculated features and corrosion inhibition.

CORROSION INHIBITORS FOR MAGNESIUM ALLOYS

Magnesium and its alloys are light and have many uses in the automotive, aerospace and consumer goods industries. Mg is quite reactive, so corrosion protection strategies are needed to be to extend the service life of Mg structural elements. As with aluminium and other metals, the vast number of small organic molecules that are synthetically accessible can also provide an almost infinite number of possibilities for inhibitors that are optimized for Mg corrosion inhibition. Advances in this field can also benefit the new battery technologies that play such an important part in our highly technological world. Small organic modulators of metal dissolution play an important role in the successful commercial development of new battery technologies.

Our research on Mg corrosion inhibitors used data from medium throughput experiments in which corrosion inhibition was assessed by the degree of hydrogen evolution from a standard metal aliquot [8]. The data on inhibition efficiencies (IE) of >150 compounds on nine different magnesium-based materials (six Mg alloys and three grades of pure Mg) were published recently [8]. All inorganic substances, molecules with molecular weights > 350 Da, and those containing under-represented molecular features (e.g., phenylphosphoric acid) or acid salts were excluded from the model. The 71 remaining molecules were used to train models describing the relationship between molecular properties and the corrosion inhibition. The dataset was divided in a training (85%) and test set (15%).

The molecules were encoded by molecular descriptors or features.  These were properties from in vacuo DFT calculations (performed at the TPSSh/def2SVP level of theory using Turbomole 7.2) and three simple chemical features: the number of OH groups, number of C=O groups and a one-hot descriptor denoting aromatic versus aliphatic character. Conspicuously, unlike similar DFT studies on Al alloys, there was a very significant correlation between corrosion inhibition and HOMO-LUMO gap for Mg and alloys (r~0.7). Thus, it appears likely that the relevance of frontier orbital features is metal dependent. The corrosion inhibition models were therefore trained on these four molecular features.

Two Mg samples were analysed, Mg with 220 ppm Fe (CP Mg 220) and 342 ppm Fe (CP Mg 342). The neural network model for CP Mg 220 could recapitulate the corrosion inhibition of molecules in the test set with an r2 of 0.89 and RMSE of 25% (Figure 3). Note that the data sets contained corrosion accelerants (higher H2 evolution than control) as well as inhibitors (lower H2 evolution than control).

All data were used to train an inhibition model making a priori predictions of seven organic agents whose inhibition had not been measured. These compounds fell within or close to the domain of applicability of the model. As Table 1 shows, the model made useful predictions of these new potential inhibitors, with the r2 for predicted versus measured inhibitions of 0.74 and RMSE of 33%.

CP Mg 342 corrodes more slowly than CP Mg 220. There was a strong correlation between the IEs for the two Mg materials (r2 value of 0.87), so the CP Mg 220 neural network model would predict similar order of inhibition for new compounds for CP Mg 342.

RECENT AND FUTURE TECHNOLOGIES FOR DESIGN OF CORROSION CONTROL AGENTS.

Additive manufacturing (3D) printing is now possible for diverse types of metals and alloys. There is potential for these techniques to generate large numbers of metals and allows in a format allowing large scale assessment of inhibition by chemical libraries. These methods can also print gradients, where one end of a sample is pure metal of one type and the other end pure metal of another type, with intermediate regions presenting different mixtures of the metals in the alloys. However, it would also be important to assess whether 3D printed samples corrode in the same way and rate as metals and alloys prepared in bulk by commercial processes. In this context, gradients incomposition are not common in industrial practice with many alloying elements ending up in discrete impurity or intermetallic particles.

Deep learning methods are being developed at a dizzying pace, keeping up with the literature in this field is challenging. One of the main advantages of deep learning methods (large, complex networks with multiple internal layers consisting of many neurons) over earlier shallow learning methods (three-layer networks containing a small number of neurons) is that they generate very effective latent features from simple representations of molecules such as molecular graphs or text strings (e.g., SMILES). Convolutional neural networks are very good at generating these latent features and using them to model quantitative relationships between molecules and inhibition, and a recent variant, the graph convolutional neural network is particularly well suited to molecules. However, deep learning methods require very large data sets for training so cannot be used effectively for design and discovery of new corrosion control agents.  However, the potential of deep neural networks and autoencoders to model corrosion inhibitors was foreshadowed by Schiessler and colleagues in a recent paper.[20]

The recent development of rapid and accurate quantum machine learning methods now allows accurate properties of molecules to be generated orders of magnitude faster than first principles quantum chemical or density functional theory methods [21, 22].  This should provide better features for training models, and the means to study the interactions of inhibitors with a metal surface using model systems that are more realistic (include solvents, ions, mixed surface compositions etc.) and account for chemical reactions.

New encoder-decoder networks and generative machine learning methods have solved a long-standing issue with quantitative structure-property relationships modelling, how do you get a trained model to predict new chemical with improved properties directly? Finally, the availability of massive ‘make on demand’ chemical libraries such as the 30+ billion ZINC-22 library have now greatly simplified the process of obtaining new molecules predicted to be useful by machine learning models [23]. In many cases, these are available from large chemical libraries or can be made on demand, minimizing or eliminating the need for in-house synthesis. The first experiments in autonomous chemical discovery capabilities with no human in the loop, incorporating generative models or other types of AI methods such as evolutionary algorithms, promise the ability to ‘evolve’ molecules towards improved properties (fitness’s) such as high inhibition, coating compatibility, cost, toxicity etc.[24]

CONCLUSIONS

Increasingly, we and others have demonstrated how machine learning models of the corrosion inhibition behaviour of diverse small organic molecules can accelerate the design of discovery of more benign replacements for current toxic materials. These exciting and broadly applicable technologies represent a paradigm shift in rapid discovery of novel, effective corrosion inhibitors for a range of metals and alloys, in diverse environments. It is important to concentrate research effort on developing high throughput methods for generating corrosion inhibition data as this is the primary reason for corrosion control missing out on the clear benefits of machine learning methods. 9

AUTHOR DETAILS

Photos are to be submitted in 600 DPI. One of the authors should be nominated as the corresponding author in case of feedback or any other communication required between the author and the Conference. The corresponding author and contact details should also be clearly stated in list of authors at the start of the paper.

1. Winkler is corresponding author of this paper. He is Professor of Biochemistry and Chemistry at La Trobe University, Professor of Pharmacy at the University of Nottingham, and Professor of Medicinal Chemistry at Monash University. He applies computational chemistry, AI, and machine learning to the design of drugs, materials, and green corrosion inhibitors. Hewas awarded an ACS Skolnik award and AMMA Medal. He has written >300 research articles and is an inventor on 25 patents.
2. Hughes is an Adjunct Professor at Deakin University’s Institute for Frontier Materials and CSIRO Fellow at CSIRO Minerals. He has over 40 years of experience in characterisation in materials science (catalysis and surface science, solid oxide ceramics, coatings development, corrosion and inhibition Al-alloys). His current research interests include novel physicochemical characterisation techniques applied to microstructure and corrosion research.
3. Mol is Professor Corrosion Technology and Electrochemistry at Delft University of Technology, and he is Scientific Director of the 4TU.High-Tech Materials Centre in the Netherlands. The specific research focus areas of Molembrace local corrosion, surface treatment, adhesive bonding, greeninhibitors, active protective and self-healing coatings. Since 2017 he is Editor[1]in-Chief of Elsevier’s Corrosion Science and has served the European Federation of Corrosion in various leading positions. He received the European Corrosion Medal in 2022. .
4. Özkan is a PhD candidate at Delft University of Technology and also serves as the Scientific Communication Officer board member for the young branch of the European Federation of Corrosion. His primary research focus is to enhance the understanding of corrosion inhibition and self-healing mechanisms within active protective aerospace coatings. His research interests lie at the intersection of electrochemistry, spectroscopy, and machine learning. 10

Würger is a postdoctoral researcher at the Institute of Surface Science at Helmholtz-Zentrum Hereon. His major research interest is located at the interface of computational chemistry, machine learning and software development. His research activities are focused on a deeper understanding of underlying corrosion mechanisms of magnesium using atomistic simulations. Moreover, he develops quantitative-structure property relationship models to identify effective dissolution modulators for magnesium-based applications ranging from battery to engineering materials.

Feiler is a scientist at the Institute of Surface Science at Helmholtz-Zentrum Hereon and Deputy Head of the Department of Interface Modelling. His major research interest is located at the interface of materials informatics and atomistic simulations. Among his activities are the application of (un)supervised machine learning approaches to identify effective, benign dissolution modulators for light engineering metals and to predict how materials and structures behave during their service-life.

1. Lamaka specializes in understanding metal degradation processes. In particular, she studies the ways of controlling degradation of magnesium for the purpose of structural engineering applications (automotive, aerospace and consumer goods), primary Mg-air aqueous batteries and bioabsorbable medical implants. She has co-authored more than 90 ISI papers with h-index 32 and more than 4000 citations, 6 book chapters, 3 granted patents and over 50 talks of which 12 were invited or keynotes. Over the last 10 years, the focus of her work was on corrosion inhibition of Mg and Al alloys, development of pre-treatments and protective self-healing coatings for Mg and Al. She is also active in the field of high-throughput experimental testing of corrosion inhibitors and combining it with in silico, computational, screening of corrosion inhibitors by using the machine learning approach. Lamaka also develops experimental tools for localized electrochemical measurements in corrosion studies. 11

REFERENCES

[1] Le T, Epa VC, Burden FR, Winkler DA. Quantitative Structure-Property Relationship Modeling of Diverse Materials Properties. Chem Rev. 2012;112:2889-919.

[2] Muratov EN, Bajorath J, Sheridan RP, Tetko IV, Filimonov D, Poroikov V, et al. QSAR without borders. Chem. Soc. Rev 2020;49:3525-64.

[3] Garcia SJ, Muster TH, Ozkanat O, Sherman N, Hughes AE, Terryn H, et al. The influence of pH on corrosion inhibitor selection for 2024-T3 aluminium alloy assessed by high-throughput multielectrode and potentiodynamic testing. Electrochim Acta. 2010;55:2457-65.

[4] Muster TH, Hughes AE, Furman SA, Harvey T, Sherman N, Hardin S, et al. A rapid screening multi-electrode method for the evaluation of corrosion inhibitors. Electrochim Acta. 2009;54:3402-11.

[5] Breedon M, Collis G, Best AS. Faster High Throughput Electrochemical Testing for Batteries. ECS Meet Abstr. 2019;MA2019-02:613.

[6] White PA, Smith GB, Harvey TG, Corrigan PA, Glenn MA, Lau D, et al. A new high-throughput method for corrosion testing. Corros Sci. 2012;58:327-31.

[7] Shi P, Li B, Huo J, Wen L. A Smart High-Throughput Experiment Platform for Materials Corrosion Study. Sci Program. 2016;2016:6876241.

[8] Lamaka SV, Vaghefinazari B, Mei D, Petrauskas RP, Höche D, Zheludkevich ML. Comprehensive screening of Mg corrosion inhibitors. Corros Sci. 2017;128:224-40.

[9] Feiler C, Mei D, Vaghefinazari B, Wurger T, Meissner RH, Luthringer-Feyerabend BJC, et al. In silico screening of modulators of magnesium dissolution. Corros Sci. 2020;163:108245.

[10] Hughes AE, Winkler DA, Carr J, Lee P, Yang Y, Laleh M, et al. Corrosion Inhibition, Inhibitor Environments, and the Role of Machine Learning. Corros Mate Degrad. 2022;3:672-93.

[11] Winkler DA, Breedon M, Hughes AE, Burden FR, Barnard AS, Harvey TG, et al. Towards chromate-free corrosion inhibitors: structure-property models for organic alternatives. Green Chem. 2014;16:3349-57.

[12] Winkler DA, Breedon M, White P, Hughes AE, Sapper ED, Cole I. Using high throughput experimental data and in silico models to discover alternatives to toxic chromate corrosion inhibitors. Corros Sci. 2016;106:229-35.

[13] Würger T, Mei D, Vaghefinazari B, Winkler DA, Lamaka SV, Zheludkevich ML, et al. Exploring structure-property relationships in magnesium dissolution modulators. npj Mater Degrad. 2021;5:2.

[14] Würger T, Wang L, Snihirova D, Deng M, Lamaka S, Winkler DA, et al. Data-driven Selection of Electrolyte Additives for Aqueous Magnesium Batteries. J Mater Chem A. 2022.

[15] Winkler DA. Predicting the Performance of Organic Corrosion Inhibitors. Metals. 2017;7.

[16] Galvao TLP, Novell-Leruth G, Kuznetsova A, Tedim J, Gomes JRB. Elucidating Structure-Property Relationships in Aluminum Alloy Corrosion Inhibitors by Machine Learning. J Phys Chem C. 2020;124:5624-35.

[17] Dibetsoe M, Olasunkanmi LO, Fayemi OE, Yesudass S, Ramaganthan B, Bahadur I, et al. Some Phthalocyanine and Naphthalocyanine Derivatives as Corrosion Inhibitors for Aluminium in Acidic Medium: Experimental, Quantum Chemical Calculations, QSAR Studies and Synergistic Effect of Iodide Ions. Molecules. 2015;20:15701-34.

[18] Harvey TG, Hardin SG, Hughes AE, Muster TH, White PA, Markley TA, et al. The effect of inhibitor structure on the corrosion of AA2024 and AA7075. Corros Sci. 2011;53:2184-90.

[19] Atlam FM, Mohamed AA, Awad MK. New Theoretical Insights about Anticorrosive Effects and Adsorption Mechanism of Some alpha-Amino Acids on Al Surface: DFT, MEP, FMO, NBO, QSAR, Fukui Functions and Monte Carlo Simulation. Protect Metals Phys Chem Surf. 2022;58:1054-70.

[20] Schiessler EJ, Wurger T, Lamaka SV, Meissner RH, Cyron CJ, Zheludkevich ML, et al. Predicting the inhibition efficiencies of magnesium dissolution modulators using sparse machine learning models. npj Comp. Mater. 2021;7.

[21] Smith JS, Isayev O, Roitberg AE. ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost. Chem. Sci. 2017;8:3192-203.

[22] Smith JS, Nebgen BT, Zubatyuk R, Lubbers N, Devereux C, Barros K, et al. Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning. Nat Comm. 2019;10:2903.

[23] Tingle BI, Tang KG, Castanon M, Gutierrez JJ, Khurelbaatar M, Dandarchuluun C, et al. ZINC-22─A Free Multi[1]Billion-Scale Database of Tangible Compounds for Ligand Discovery. J Chem Inf Model. 2023;63:1166-76.

[24] Le TC, Winkler DA. Discovery and Optimization of Materials Using Evolutionary Approaches. Chem. Rev. 2016;116:6107-32.