metals

Review
Predicting the Performance of Organic
Corrosion Inhibitors
David A. Winkler 1,2,3,4,5 ID

1 La Trobe Institute for Molecular Science, La Trobe University, Kingsbury Drive, Bundoora 3086, Australia;
[email protected]; Tel.: +61-466-258-600
2 Monash Institute of Pharmaceutical Sciences, Monash University, 392 Royal Parade, Parkville 3052, Australia
3 School of Chemical and Physical Sciences, Flinders University, Bedford Park 5046, Australia
4 CSIRO Manufacturing, Bayview Avenue, Clayton 3168, Australia
5 School of Pharmacy, University of Nottingham, Nottingham NG7 2RD, UK

Received: 20 November 2017; Accepted: 5 December 2017; Published: 8 December 2017

Abstract: The withdrawal of effective but toxic corrosion inhibitors has provided an impetus for the
discovery of new, benign organic compounds to fill that role. Concurrently, developments in the
high-throughput synthesis of organic compounds, the establishment of large libraries of available
chemicals, accelerated corrosion inhibition testing technologies, and the increased capability of
machine learning methods have made discovery of new corrosion inhibitors much faster and cheaper
than it used to be. We summarize these technical developments in the corrosion inhibition field
and describe how data-driven machine learning methods can generate models linking molecular
properties to corrosion inhibition that can be used to predict the performance of materials not
yet synthesized or tested. We briefly summarize the literature on quantitative structure–property
relationships models of small organic molecule corrosion inhibitors. The success of these models
provides a paradigm for rapid discovery of novel, effective corrosion inhibitors for a range of metals
and alloys in diverse environments.

Keywords: corrosion inhibitors; high-throughput corrosion inhibition testing; machine learning;

quantitative structure–property relationships (QSPR); organic molecules; molecular design;
data-driven models

1. Introduction
Corrosion is responsible for an excessive amount of catastrophic failure in many different industries,
causing death, injury, and capital loss. Corrosion prevention and treatment has thus become a
multi-trillion-dollar imposition on industry. Corrosion may be inhibited by a range of surface treatments
and coatings, many of which are highly effective. For example, in the aerospace industries, chromates
have been the mainstay of corrosion treatment of alloys used in aircraft. However, health and safety
and environmental concerns have led to the ban or restricted use of corrosion inhibitors containing such
elements as tin, chromium, and lead. For example, the use of chromate inhibitors is being phased out in
most countries. It has been estimated that workers exposed to chromate residues when aircraft is sanded
and recoated have up to a 250,000-fold higher risk of cancer than the general public [1].
Consequently, there has been a large research effort to generate inhibitors and coatings that are
very effective and more benign to workers, the public, and the environment. Small organic compounds
show promise as corrosion inhibitors, although their mechanisms of action are far from clear.
Recent developments in automated high-throughput chemical synthesis [2,3], the availability of
large libraries of organic compounds, high-throughput methods of assessing corrosion inhibition by
mass loss, electrochemistry or other means, and the rapid growth in the capabilities of machine learning
methods such as deep learning provide an unprecedented opportunity to discover or design new,

Metals 2017, 7, 553; doi:10.3390/met7120553 www.mdpi.com/journal/metals

Metals 2017, 7, 553 2 of 8

more effective corrosion inhibitors [4,5]. Small molecule, synthetically accessible space is also vast,
estimated to be ~10100 , providing rich opportunities for discovery of novel, effective small molecule
corrosion inhibitors [6]. However, the size of chemistry space also means that high-throughput
experimentation cannot explore a significant fraction of it alone, unless other data-driven methods like
design of experiments [7], structure–property relationship modelling, and evolutionary methods are
used to extensively leverage the data that high throughput screening (HTS) methods can generate.
This paper briefly reviews some of the more important developments in the high-throughput
synthesis, design, and modelling of organic corrosion inhibitors. The main focus of the review is the use
of the quantitative structure–property relationship (QSPR) method to predict the corrosion inhibitory
properties of organic compounds [8–10]. Such models can be used to understand the relationships
between the chemical structure of inhibitors and their efficacy and to allow the inhibition of compounds
not yet synthesized or tested to be predicted. They can also be used as surrogate fitness functions for
the evolutionary design of new corrosion inhibitors with multiple desirable properties [11–13].

2. High-Throughput Synthesis and Testing of Organic Corrosion Inhibitors

Although high-throughput and combinatorial synthesis of small organic compounds is well
established in the pharmaceutical industry, it is rarely used in corrosion inhibition research. This is
likely due to the lack of high-throughput corrosion inhibition testing methods, and the existence of
very large libraries of commercially available organic compounds that can be tested for corrosion
inhibition. Development of high-throughput, direct, or surrogate methods of assessing corrosion
inhibition performance is therefore critical to progress in this field.
Significant progress has been made recently in high-throughput corrosion inhibition testing.
The most important factor is how the fast, surrogate testing methods used in high-throughput testing
correlate with “real world” corrosion inhibition. Currently, electrochemical methods, mass loss,
and photometric methods have been employed to assess corrosion inhibition. Seminal work has
been reported by Chambers et al., who used direct current polarization between two aerospace
aluminium alloy (AA2024) wire electrodes and a multiple-electrode testing system to assess the
corrosion inhibition of fifty chemistries in just 9 h. The results correlated highly with those of
extended testing over 10 days [14]. This research team extended their work by scoring corrosion using
fluorometric detection of Al3+ concentrations [15]. They measured corrosion inhibition of 14 corrosion
inhibitors over a wide range of initial pH values for a period between 1 and 7 days [16]. Their later
work developed a system for rapidly assessing inhibition characteristics of 100 separate chemistries
using direct current (DC) polarization, cyclic voltammetry of re-deposited copper, and fluorometric
detection of Al3+ [17,18]. Kallip and coworkers reported a novel scanning vibrating electrode technique
for assessing corrosion inhibitors [19]. They were able to accurately determine percent corrosion
inhibition efficiencies of Fe and Zn for four inhibitors. He et al. described a high-throughput
electrochemical impedance spectroscopy (HT-EIS) method for the rapid evaluation of corrosion
coatings. They developed a 12-element, spatially addressable electrochemical platform interfaced to a
commercial EIS instrument [20]. Recently, White et al. described a novel method for assessing corrosion
inhibition via a high-throughput testing rig (Figure 1) [1,21]. Up to 88 simultaneous corrosion inhibition
tests can be carried out on a single plate, with positive and negative controls, in approximately one day.
Corrosion inhibition was assessed using a novel, robust computerized image processing method.
More recently, Shi et al. reported a similar automated system for corrosion assessment using optical
imaging that also showed a linear relationship between an apparent grey scale value of the image and
the depth of corrosive pitting in the specimen [22].
Metals 2017, 7, 553 3 of 8
Metals 2017, 7, 553 3 of 8

Figure 1. High-throughput corrosion inhibition rig consisting of a 10-mm-thick polycarbonate

Figure 1. High-throughput corrosion inhibition rig consisting of a 10-mm-thick polycarbonate clamped
clamped to a 10-mm-thick block of polydimethylsiloxane rubber, an abraded plate of alloy, and a 5-
to a 10-mm-thick block of polydimethylsiloxane rubber, an abraded plate of alloy, and a 5-mm-thick
mm-thick metal baseplate. The polycarbonate and polydimethylsiloxane sheets have an 8 × 11 grid of
metal baseplate. The polycarbonate and polydimethylsiloxane sheets have a 8 × 11 grid of holes for
holes
test for test solutions
solutions 6 mm in diameter.
6 mm in diameter. Used withUsed with permission
permission frometWinkler
from Winkler al. [1]. et al. [1].

3. Machine Learning Modelling Methods

3. Machine Learning Modelling Methods
Clearly, the structures of organic compounds play a major role in how effectively they inhibit
Clearly, the structures of organic compounds play a major role in how effectively they inhibit
metal corrosion. In essence, changes to chemical structure directly modulate corrosion inhibition, a
metal corrosion. In essence, changes to chemical structure directly modulate corrosion inhibition,
pattern recognition problem. The pharmaceutical industry has a long history of developing methods
a pattern recognition problem. The pharmaceutical industry has a long history of developing methods
for generating quantitative models describing similar effects (on biological targets in this case).
for generating quantitative models describing similar effects (on biological targets in this case).
Recently, this large body of knowledge has been leveraged by materials science to design and
Recently, this large body of knowledge has been leveraged by materials science to design and discover
discover new bespoke materials with hitherto inaccessible properties. As these methods are data-
new bespoke materials with hitherto inaccessible properties. As these methods are data-driven,
driven, the emergence of reliable high-throughput methods for screening potential corrosion
the emergence of reliable high-throughput methods for screening potential corrosion inhibitors
inhibitors described above means that quantitative modelling methods will be used more frequently
described above means that quantitative modelling methods will be used more frequently in corrosion
in corrosion inhibition research.
inhibition research.
Quantitative models linking chemical structure to properties can be developed using machine
Quantitative models linking chemical structure to properties can be developed using machine
learning or other statistical methods. These methods—quantitative structure–activity relationships
learning or other statistical methods. These methods—quantitative structure–activity relationships
(QSAR) and quantitative structure–property relationships (QSPR) modelling—are essentially the
(QSAR) and quantitative structure–property relationships (QSPR) modelling—are essentially the same.
same. The main difference is in the types of properties being modelled (biological versus non-
The main difference is in the types of properties being modelled (biological versus non-biological)
biological) and the ways in which the molecular or material is represented mathematically in the
and the ways in which the molecular or material is represented mathematically in the modelling
modelling process [8,10,23].
process [8,10,23].
QSAR and QSPR methods involve several processes [24,25]. A set of molecules or materials are
QSAR and QSPR methods involve several processes [24,25]. A set of molecules or materials are
collected and tested for the desired property. The structural and physicochemical properties of the
collected and tested for the desired property. The structural and physicochemical properties of the
molecule or material are encoded mathematically (these are called molecular descriptors) in a way
molecule or material are encoded mathematically (these are called molecular descriptors) in a way
this is relevant to the property being modelled, and a subset selected in a context-dependent way. A
this is relevant to the property being modelled, and a subset selected in a context-dependent way.
mathematical relationship, often nonlinear, is found between the descriptors and the property being
A mathematical relationship, often nonlinear, is found between the descriptors and the property being
modelled, and the predictive power of the model assessed using an independent test set of
modelled, and the predictive power of the model assessed using an independent test set of compounds
compounds not used to generate the model [26]. Finally, the model can be interrogated to understand
not used to generate the model [26]. Finally, the model can be interrogated to understand what features
what features of the molecules improve or degrade their performance, and the model can be used to
of the molecules improve or degrade their performance, and the model can be used to predict the
predict the properties of compounds not yet synthesized or tested. Very large real or virtual libraries
properties of compounds not yet synthesized or tested. Very large real or virtual libraries of organic
of organic compounds can be screened very quickly this way, so long as the screened molecules lie
compounds can be screened very quickly this way, so long as the screened molecules lie close to the
close to the domain of applicability of the models (the region of chemical and property space in which
domain of applicability of the models (the region of chemical and property space in which the model
the model was trained).
was trained).
Mathematical relationships between the selected descriptors and the property being modelled
Mathematical relationships between the selected descriptors and the property being modelled can
can be generated by linear (e.g., multiple linear regression) or nonlinear (e.g., neural network,
be generated by linear (e.g., multiple linear regression) or nonlinear (e.g., neural network, recursive
recursive partitioning, kernel or polynomial regression) methods. Neural network and other machine
partitioning, kernel or polynomial regression) methods. Neural network and other machine learning
learning methods can generate models quickly and effectively and require few or no assumptions to
be made about the form of the mathematical relationship, as they are universal approximators [27,28].
Metals 2017, 7, 553 4 of 8

methods can generate models quickly and effectively and require few or no assumptions to be made
about the form of the mathematical relationship, as they are universal approximators [27,28]. The rise
of deep learning algorithms within the last five years has stimulated the use of neural network and
machine learning approaches substantially [4,5,29].

4. Computational Models of Corrosion Inhibitory Properties of Organic Compounds

There have been relatively few reports of QSPR models for corrosion inhibition by small organic
molecules, and fewer still that are robust. Many early reported models were based on very small
data sets with limited chemical diversity. This increased the probability of chance correlations that
were not indicative of causative relationships between the compounds and the inhibition. A relatively
large number of studies with small number of inhibitors reported a relationship between the frontier
orbital properties of small organic molecule (principally the energies of the highest occupied molecular
orbital (HOMO) and lowest unoccupied molecular orbital (LUMO): EHOMO , ELUMO ; and band gap
energies ELUMO − EHOMO ), calculated by various quantum mechanical (QM) methods and corrosion
inhibition. For example, Sastri and Perumareddi reported correlations of corrosion rates with EHOMO ,
the energy gap, and Hammett’s parameter (σ) for a small set of organic compounds [30]. Subsequently,
Ozcan and Dehri reported a similar correlation between frontier orbital properties of three organic
inhibitors, thioacetamide, thiourea, and thiobenzamide [31]. A second study by these authors reported
apparent correlations between corrosion inhibition and frontier orbital properties for three other
organic compounds—thiourea, methylthiourea, and phenylthiourea—with similar structures and high
corrosion inhibition (93%, 96%, and 97%, respectively). The small changes in structure and the small
differences inhibition values relative to experimental error suggest that these correlations were not
statistically valid [32]. Sastri et al. subsequently reported that corrosion inhibition correlated with
EHOMO , the HOMO–LUMO gap, the chemical softness of the inhibitor, chemical potential, and the
fraction of charge transferred from the inhibitor to the metal [33]. However, this result was based on few
compounds and, like all of the above studies, did not include allowance for solvent effects, ionization,
and speciation at testing pH. Recently, Bedair reported a quantum chemical study of corrosion
inhibition of pyridine, quinolone, acridine, and their n-hexadecyl derivatives on steel. He used
density functional theory (DFT), ab initio calculations, and semi-empirical methods to find correlations
between molecular structure and corrosion inhibition efficiency. He reported that calculated dipole
moments, EHOMO , ELUMO , and the HOMO–LUMO gap correlated with corrosion inhibition, although
the small number of compounds used and their similarities in structure throw doubt on the utility of
these correlations [34]. Saha et al. similarly investigated 2-aminopyrazine, 2-amino-5-bromopyrazine,
3-amino pyrazine-2-thiol, and 3-amino-6-bromopyrazine-2-thiol as corrosion inhibitors using density
functional theory (DFT) and molecular dynamics (MD) simulations [35]. In these studies, a solvent
model, conductor-like screening model (COSMO), and neutral and protonated species were used in
the correlations. However, only qualitative comparisons between corrosion inhibition and quantum
chemical parameters were found. In all of the above studies the number of organic compound studies
were too few to allow a test set to be used to quantify the predictivity of the models.
Considerable doubt has been thrown on the usefulness of molecular properties calculated by
quantum chemical methods by two recent studies that used a much larger and chemically diverse set
of organic corrosion inhibitors. Winkler et al. showed that robust, predictive models linking computed
molecular properties to corrosion inhibition can be generated using neural networks [36]. However,
the molecular properties that most influenced the inhibition were quite arcane and hard to interpret.
The first study used mass loss as a measure of corrosion of aerospace aluminium alloys and employed
28 diverse small organic molecule inhibitors. They found that, whether or not speciation was included,
there was essentially no correlation between ionization potential, HOMO or LUMO energies, or any
other quantum chemically derived descriptors and corrosion efficiency. A subsequent and larger
study using 100 diverse inhibitors and a novel high-throughput corrosion inhibition experimental
methods also showed negligible correlation between quantum chemically derived parameters and
Metals 2017, 7, 553 5 of 8

corrosion inhibition. However, in this case, good, robust models were developed that linked molecular
Metals 2017, 7, 553 5 of 8
properties calculated by non-quantum chemical methods and corrosion inhibition. These models could
make reliable,
corrosion quantitative
inhibition. Thesepredictions of corrosion
models could inhibition
make reliable, for compounds
quantitative not usedoftocorrosion
predictions train the
model (Figure 2) [1].
inhibition for compounds not used to train the model (Figure 2) [1].

Figure 2.2.Quantitative
Figure Quantitative prediction
prediction of corrosion
of corrosion inhibitory
inhibitory properties
properties of compounds
of organic organic compounds for
for aerospace
aerospace alloys AA-2024 generated using molecular descriptors (not
alloys AA-2024 generated using molecular descriptors (not derived from quantum mechanicalderived from quantum
mechanical calculations).
calculations). Inhibition isInhibition is scaled
scaled between between
0 (no 0 (no and
inhibition) inhibition) and inhibition).
10 (highest 10 (highest inhibition).
The circles
The circles represent the performance of the model in predicting training
represent the performance of the model in predicting training set, and the triangles set, and represent
the triangles
the
represent theof
performance performance
the model in ofpredicting
the model the
in predicting the test
test set. Used withset. Used withfrom
permission permission
Winklerfrom
et al.Winkler
[1].
et al. [1].

There is potential for more sophisticated quantum chemical methods to contribute to modelling of
There is potential for more sophisticated quantum chemical methods to contribute to modelling
corrosion inhibition. Recently, Breedon et al. reported a novel 3D-QSPR method—comparative molecular
of corrosion inhibition. Recently, Breedon et al. reported a novel 3D-QSPR method—comparative
surface analysis (CoMSA)—that employed 3D distributions of electronegativity, polarizability, and van der
molecular surface analysis (CoMSA)—that employed 3D distributions of electronegativity,
Waals volume on the molecular surfaces of 28 small organic molecules as descriptors [37]. This method
polarizability, and van der Waals volume on the molecular surfaces of 28 small organic molecules as
could make a qualitative prediction of corrosion inhibition efficiency, identifying high-performing
descriptors [37]. This method could make a qualitative prediction of corrosion inhibition efficiency,
corrosion inhibitor candidates. This approach may be elaborated to improve the accuracy of prediction,
identifying high-performing corrosion inhibitor candidates. This approach may be elaborated to
take into account solvents and speciation, and allow quantitative, or at least semi-quantitative predictions
improve the accuracy of prediction, take into account solvents and speciation, and allow quantitative,
to be made about corrosion inhibition.
or at least semi-quantitative predictions to be made about corrosion inhibition.
When other types of measured or computed molecular descriptors are used, it is possible
When other types of measured or computed molecular descriptors are used, it is possible to
to build robust and predictive models of corrosion inhibition from large data sets derived from
build robust and predictive models of corrosion inhibition from large data sets derived from high-
high-throughput testing of large chemical libraries. These descriptors relate to the types of atoms and
throughput testing of large chemical libraries. These descriptors relate to the types of atoms and their
their valence in compounds, chemical graph properties, physicochemical properties such as dipole
valence in compounds, chemical graph properties, physicochemical properties such as dipole
moment and lipophilicity (polarity), and many other molecular characteristics. More than 3000 of these
moment and lipophilicity (polarity), and many other molecular characteristics. More than 3000 of
mathematical representations of molecules can be computed using commercially available packages
these mathematical representations of molecules can be computed using commercially available
such as Dragon [38]. One of the first studies of the structure–activity relationships in 400 organic
packages such as Dragon [38]. One of the first studies of the structure–activity relationships in 400
compounds of corrosion inhibitors was reported by Horner and Meisel in 1978 [39]. Subsequently,
organic compounds of corrosion inhibitors was reported by Horner and Meisel in 1978 [39].
corrosion inhibition for iron and nickel in acidic solution was measured for four organic inhibitors
Subsequently, corrosion inhibition for iron and nickel in acidic solution was measured for four
by Jayalakshmi and Muralidharan [40]. They generated qualitative structure–property relationships,
organic inhibitors by Jayalakshmi and Muralidharan [40]. They generated qualitative structure–
showing that the concentration of the inhibitor, hydrophobicity, and π electron density played a role in
property relationships, showing that the concentration of the inhibitor, hydrophobicity, and π
inhibitor efficiency.
electron density played a role in inhibitor efficiency.
Recently, Keshavarz et al. [41] reported a simple method for predicting corrosion inhibition
Recently, Keshavarz et al. [41] reported a simple method for predicting corrosion inhibition
efficiency for steel of small organic molecules, principally imidazoles and benzimidazoles. They studied
efficiency for steel of small organic molecules, principally imidazoles and benzimidazoles. They
34 diverse chemicals and derived a simple linear regression model for inhibition:
studied 34 diverse chemicals and derived a simple linear regression model for inhibition:
η =η 38.47
= 38.47
+ 20.21n(N) − −7.98n(O
+ 20.21n(N) 7.98n(O++NH 14.94η++ −
NH2 )2)++14.94η 17.93η−−
− 17.93η (1)
(1)
where η is the corrosion inhibition efficiency, n(N) is the number of nitrogen atoms; n(O + NH2) is the
sum of the number oxygen atoms and amino groups; η+ and η− are the positive and negative effects
of structural parameters on the inhibition efficiency. Their model was more accurate in predicting the
corrosion efficiency of compounds in a test set of 11 compounds not used to generate the model than
Metals 2017, 7, 553 6 of 8

where η is the corrosion inhibition efficiency; n(N) is the number of nitrogen atoms; n(O + NH2 ) is the
sum of the number oxygen atoms and amino groups; η+ and η− are the positive and negative effects
of structural parameters on the inhibition efficiency. Their model was more accurate in predicting the
corrosion efficiency of compounds in a test set of 11 compounds not used to generate the model than
two QSPR models reported by Zhang et al. [42] derived from quantum chemically derived parameters.
However, the range of inhibition values for the training and test sets was not high (60–95% for the
test set) relative to the prediction errors, and the way in which the η+ and η− parameters correct the
predictions of inhibition is not clear.
The modelling study reported by Zhang et al. involved 34 imidazole and benzimidazoles corrosion
inhibitors and used quantum chemically derived descriptors augmented by topographical descriptors
called Molecular Connectivity Indices (MCI) [42]. They stated that a previous paper by Zhao et al.
(which they did not reference) could not find strong QSPR models for organic corrosion inhibitors
using quantum chemical descriptors alone. They augmented the QM descriptors by MCI descriptors
and dramatically improved the models. This provided the rationale for the use of MCI descriptors in
Zhang et al.’s study. Using this composite descriptor set they could generate models of the relationship
between molecular properties and corrosion inhibition. Inhibition was assessed by weight loss. The QM
descriptors used included EHOMO , ELUMO , partial charges, electron densities, frontier orbital properties,
and polarizability. These were augmented by log octanol-water partition coefficients (a measure of
lipid solubility), nX and δi ’ topological indices derived from the chemical graph of each inhibitor.
The data set of 34 inhibitors used 16 of the 18 descriptors, making overfitting likely unless care is taken.
Stepwise linear regression generated two models with six parameters that reproduced the training
set of 34 compounds with r2 = 0.81 and a standard error of 10%. One model included the effect of
electron-donating and -accepting properties of a metal surface. Unfortunately, these authors did not
use an independent test set to test the predictivity of the model for new data.

5. Conclusions and Perspective

The need for new, safe and environmentally benign corrosion inhibitors has been strengthened by
the unacceptable toxicity of existing, albeit highly effective, methods for corrosion control. There is a
very fortunate juxtaposition of this need with the large increase in capabilities of technologies relevant
to the design of new organic corrosion inhibitors. Automated, high-throughput methods of screening
large numbers of potential inhibitors will accelerate the discovery of new inhibitors and ultimately
allow for inhibitors with multiple functions (e.g., corrosion inhibition and the inhibition of methane
hydrates in undersea gas pipelines [43]).
These automated corrosion inhibition testing methods can potentially generate large data sets for
a diverse range of organic chemotypes that are very well matched to analysis by machine learning
and other statistical modelling methods. Our review shows that it is possible to generate robust
and predictive computational models that make accurate, quantitative predictions of the degree of
corrosion inhibition for organic compounds not yet tested. The most successful models reported
in the literature employ empirical molecular descriptors as models that use molecular descriptors
derived from quantum chemical calculations that are not statistically significant enough to be useful.
We anticipate that quantum chemical calculations based on model systems much closer to real world
metal/water/inhibitors systems may play an important role in the future. Although there are relatively
few reported studies of high-throughput testing and modelling of corrosion inhibition properties of
small organic molecules, we are at the bottom of the S-curve. Improvements in robotics and machine
learning will be autocatalytic, leading to a massive increase in the capabilities and reliability of methods
for the design of organic corrosion inhibitors in the short to medium term.

Conflicts of Interest: The author declares no conflict of interest.

Metals 2017, 7, 553 7 of 8

References
1. Winkler, D.A.; Breedon, M.; White, P.; Hughes, A.E.; Sapper, E.D.; Cole, I. Using high throughput
experimental data and in silico models to discover alternatives to toxic chromate corrosion inhibitors.
Corros. Sci. 2016, 106, 229–235. [CrossRef]
2. Li, J.Q.; Ballmer, S.G.; Gillis, E.P.; Fujii, S.; Schmidt, M.J.; Palazzolo, A.M.E.; Lehmann, J.W.; Morehouse, G.F.;
Burke, M.D. Synthesis of many different types of organic small molecules using one automated process.
Science 2015, 347, 1221–1226. [CrossRef] [PubMed]
3. Service, R.F. The synthesis machine. Science 2015, 347, 1190–1193. [CrossRef] [PubMed]
4. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef] [PubMed]
5. Ma, J.S.; Sheridan, R.P.; Liaw, A.; Dahl, G.E.; Svetnik, V. Deep neural nets as a method for quantitative
structure–activity relationships. J. Chem. Inf. Model. 2015, 55, 263–274. [CrossRef] [PubMed]
6. Reymond, J.L.; Awale, M. Exploring chemical space for drug discovery using the chemical universe database.
ACS Chem. Neurosci. 2012, 3, 649–657. [CrossRef] [PubMed]
7. Walters, F.H. Use of statistical mixture designs to evaluate ternary mixtures of organics as corrosion-inhibitors.
Anal. Lett. 1990, 23, 1783–1790. [CrossRef]
8. Fujita, T.; Winkler, D.A. Understanding the roles of the “two QSARs”. J. Chem. Inf. Model. 2016, 56, 269–274.
[CrossRef] [PubMed]
9. Winkler, D.A. Neural networks as robust tools in drug lead discovery and development. Mol. Biotechnol.
2004, 27, 139–167. [CrossRef]
10. Winkler, D.A.; Burden, F.R. Robust QSAR models from novel descriptors and Bayesian regularised neural
networks. Mol. Simulat. 2000, 24, 243–258. [CrossRef]
11. Le, T.C.; Winkler, D.A. A bright future for evolutionary methods in drug design. ChemMedChem 2015, 10,
1296–1300. [CrossRef] [PubMed]
12. Le, T.C.; Winkler, D.A. Discovery and optimization of materials using evolutionary approaches. Chem. Rev.
2016, 116, 6107–6132. [CrossRef] [PubMed]
13. Winkler, D.A. Biomimetic molecular design tools that learn, evolve, and adapt. Beilstein J. Org. Chem. 2017,
13, 1288–1302. [CrossRef] [PubMed]
14. Chambers, B.D.; Taylor, S.R.; Kendig, M.W. Rapid discovery of corrosion inhibitors and synergistic
combinations using high-throughput screening methods. Corrosion 2005, 61, 480–489. [CrossRef]
15. Chambers, B.D.; Taylor, S.R. The high throughput assessment of aluminium alloy corrosion using
fluorometric methods. Part I—Development of a fluorometric method to quantify aluminium ion
concentration. Corros. Sci. 2007, 49, 1584–1596. [CrossRef]
16. Chambers, B.D.; Taylor, S.R. The high throughput assessment of aluminium alloy corrosion using
fluorometric methods. Part II—A combinatorial study of corrosion inhibitors and synergistic combinations.
Corros. Sci. 2007, 49, 1597–1609. [CrossRef]
17. Taylor, S.R.; Chambers, B.D. The discovery of non-chromate corrosion inhibitors for aerospace alloys using
high-throughput screening methods. Corros. Rev. 2007, 25, 571–590. [CrossRef]
18. Taylor, S.R.; Chambers, B.D. Identification and characterization of nonchromate corrosion inhibitor synergies
using high-throughput methods. Corrosion 2008, 64, 255–270. [CrossRef]
19. Kallip, S.; Bastos, A.C.; Zheludkevich, M.L.; Ferreira, M.G.S. A multi-electrode cell for high-throughput svet
screening of corrosion inhibitors. Corros. Sci. 2010, 52, 3146–3149. [CrossRef]
20. He, J.; Bahr, J.; Chisholm, B.J.; Li, J.; Chen, Z.; Balbyshev, S.N.; Bonitz, V.; Bierwagen, G.P. Combinatorial
materials research applied to the development of new surface coatings X: A high-throughput electrochemical
impedance spectroscopy method for screening organic coatings for corrosion inhibition. J. Comb. Chem. 2008,
10, 704–713. [CrossRef] [PubMed]
21. White, P.A.; Smith, G.B.; Harvey, T.G.; Corrigan, P.A.; Glenn, M.A.; Lau, D.; Hardin, S.G.; Mardel, J.;
Markley, T.A.; Muster, T.H.; et al. A new high-throughput method for corrosion testing. Corros. Sci. 2012, 58,
327–331. [CrossRef]
22. Shi, P.; Li, B.; Huo, J.; Wen, L. A smart high-throughput experiment platform for materials corrosion study.
Sci. Program. 2016, 2016. [CrossRef]
23. Le, T.; Epa, V.C.; Burden, F.R.; Winkler, D.A. Quantitative structure–property relationship modeling of
diverse materials properties. Chem. Rev. 2012, 112, 2889–2919. [CrossRef] [PubMed]
Metals 2017, 7, 553 8 of 8

24. Engel, T.; Gasteiger, J. Chemoinformatics: Basic Concepts and Methods; Wiley VCH: Weinheim, Germany, 2018.
25. Engel, T.; Gasteiger, J. Chemoinformatics: From Methods to Applications; Wiley VCH: Weinheim, Germany, 2018.
26. Alexander, D.L.J.; Tropsha, A.; Winkler, D.A. Beware of R2 : Simple, unambiguous assessment of the
prediction accuracy of QSAR and QSPR models. J. Chem. Inf. Model. 2015, 55, 1316–1322. [CrossRef]
[PubMed]
27. Gupta, M.K.; Gupta, S.; Rawal, R.K. Impact of artificial neural networks in QSAR and computational modeling.
In Artificial Neural Network for Drug Design, Delivery and Disposition; Elsevier: Amsterdam, The Netherlands, 2016;
pp. 153–179.
28. Montanez-Godinez, N.; Martinez-Olguin, A.C.; Deeb, O.; Garduno-Juarez, R.; Ramirez-Galicia, G.
QSAR/QSPR as an application of artificial neural networks. Methods Mol. Biol. 2015, 1260, 319–333. [PubMed]
29. Winkler, D.A. Performance of deep and shallow neural networks, the universal approximation theorem,
activity cliffs, and QSAR. Mol. Inform. 2017, 36, 1600118. [CrossRef] [PubMed]
30. Sastri, V.S.; Perumareddi, J.R. Molecular orbital theoretical studies of some organic corrosion inhibitors.
Corrosion 1997, 53, 617–622. [CrossRef]
31. Ozcan, M.; Dehri, I. Electrochemical and quantum chemical studies of some sulphur-containing organic
compounds as inhibitors for the acid corrosion of mild steel. Prog. Org. Coat. 2004, 51, 181–187. [CrossRef]
32. Ozcan, M.; Dehri, I.; Erbil, M. Organic sulphur-containing compounds as corrosion inhibitors for mild steel
in acidic media: Correlation between inhibition efficiency and chemical structure. Appl. Surf. Sci. 2004, 236,
155–164. [CrossRef]
33. Sastri, V.S.; Elboujdaini, M.; Perumareddi, J.R. Utility of quantum chemical parameters in the rationalization
of corrosion inhibition efficiency of some organic inhibitors. Corrosion 2005, 61, 933–942. [CrossRef]
34. Bedair, M.A. The effect of structure parameters on the corrosion inhibition effect of some heterocyclic
nitrogen organic compounds. J. Mol. Liquids 2016, 219, 128–141. [CrossRef]
35. Saha, S.K.; Hens, A.; Murmu, N.C.; Banerjee, P. A comparative density functional theory and molecular
dynamics simulation studies of the corrosion inhibitory action of two novel N-heterocyclic organic
compounds along with a few others over steel surface. J. Mol. Liq. 2016, 215, 486–495. [CrossRef]
36. Winkler, D.A.; Breedon, M.; Hughes, A.E.; Burden, F.R.; Barnard, A.S.; Harvey, T.G.; Cole, I. Towards
chromate-free corrosion inhibitors: Structure–property models for organic alternatives. Green Chem. 2014, 16,
3349–3357. [CrossRef]
37. Fernandez, M.; Breedon, M.; Cole, I.S.; Barnard, A.S. Modeling corrosion inhibition efficacy of small
organic molecules as non-toxic chromate alternatives using comparative molecular surface analysis (comsa).
Chemosphere 2016, 160, 80–88. [CrossRef] [PubMed]
38. Mauri, A.; Consonni, V.; Pavan, M.; Todeschini, R. Dragon software: An easy approach to molecular
descriptor calculations. MATCH Commun. Math. Comput. Sci. 2006, 56, 237–248.
39. Horner, L.; Meisel, K. Corrosion-inhibitors 23 (1)—Does there exist a structure–efficiency relation in organic
inhibitors of aluminum corrosion. Werkstoffe Korros. Mater. Corros. 1978, 29, 654–664. [CrossRef]
40. Jayalakshmi, M.; Muralidharan, V.S. Correlation between structure and inhibition of organic compounds for
acid corrosion of transition metals. Indian J. Chem. Technol. 1998, 5, 16–28.
41. Keshavarz, M.H.; Esmaeilpour, K.; Golikand, A.N.; Shirazi, Z. Simple approach to predict corrosion inhibition
efficiency of imidazole and benzimidazole derivatives as well as linear organic compounds containing several
polar functional groups. Z. Anorg. Allgem. Chem. 2016, 642, 906–913. [CrossRef]
42. Zhang, S.G.; Lei, W.; Xia, M.Z.; Wang, F.Y. QSAR study on n-containing corrosion inhibitors: Quantum
chemical approach assisted by topological index. J. Mol. Struct. 2005, 732, 173–182. [CrossRef]
43. Burgazli, C.R.; Navarrete, R.C.; Mead, S.L. New dual purpose chemistry for gas hydrate and corrosion
inhibition. J. Can. Petrol. Technol. 2005, 44, 47–50. [CrossRef]

© 2017 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).