The January 2018 and February 2019 editorials in CORROSION highlighted some of the grand challenges and opportunities facing corrosion science and engineering in the 21st century and future frontiers.1-2 The challenges associated with the multi-scale, multi-physics, as well as multi-stage nature of corrosion were discussed in Parts I and II.1-2 These helped define some of the frontiers of the field. In the past, steady advances in both theory and experimental methods, as well as transformational advances in computing and computational modeling, paced progress.3 This naturally leads to the question of “what’s next” in corrosion science and corrosion control and what further leaps-in-progress are in-store? Certainly one guess might be quantum computing applied to either knowledge-centric or data-centric approaches, and another might be artificial intelligence in the realm of data-centric approaches.4
There are opportunities for big data, data analytics, and artificial intelligence to enable the next leap ahead in corrosion science that informs and guides corrosion control. “Alternative futures” may be envisioned but we shall save them for another discussion. Joining me in this most fascinating discussion is Prasanna Balachandran,(1) an author of noteworthy manuscripts in the area of accelerated and high throughput materials discovery5-12 and a “rising computational materials expert” as identified by the journal Computational Materials Science.13
Corrosion is a highly complex process involving coupled chemical, electrochemical, and solid-state reactions as noted previously.2 Both thermodynamic and kinetic factors typically control behavior and materials science, surface science, as well as bulk, interface, and occluded site (electro)chemical processes govern corrosion behavior.1-2 The different stages of even a single corrosion process such as pitting of a passivated material adds to the complexity.2,14 One of the difficulties in corrosion is the vast number of factors influencing behaviors and outcomes.15 This situation, in our opinion, is favorable for the use of data-centric approaches4 based on either aggregated data or fundamental data and work against knowledge-centric approaches based on fundamental theories and expert judgements.(2) Much theory, knowledge, and data have been acquired and curated over the last ∼100 years in corrosion, with more becoming available every day through sensors and high throughput approaches. In fact, more data have been created over the past decade than available previously since the beginning of time.
Indeed, the era of big data and data analytics has arrived and is already underway in many places (medicine, materials, and elsewhere).16-19 Schools of data science have emerged at many universities.20 Past knowledge and existing data allow us to use algorithms in order to attempt to recognize patterns identifying future corrosion performance. The hope is that these approaches are helpful for predicting not only corrosion performance but failures, too.4 Such approaches are logical as long as existing data cover a wide range of possibilities and there are no required extrapolations into sparsely populated data spaces. Rare or “unusual event” data must be factored in and understanding of rare events is generally lacking.21 Probabilistic predictive approaches such as Bayesian network models, correlative or pattern recognition approaches such as artificial neural networks are enabled but cannot forecast the surprises or rare events.22-23 In contrast, fundamental deterministic modeling of corrosion4,24-25 based on the underlying physics of all relevant processes contain predictive qualities that might apply to new circumstances lacking data or experience as long as the circumstances don’t require new or overlooked physics. If this is the case, they will fail too.
New research directions, such as active learning6,26 and physics-informed machine learning (ML),27-29 have the potential to overcome some of the short-comings and push further the promises of corrosion informatics30 (see Figure 1). However, the potential impact is yet to be fully realized. For instance, the combined ability to forecast future events with an informed knowledge about prediction uncertainties may shed new light on the scientific gaps and motivate new experiments to fill these gaps. This requires a renewed perception of the meaning of uncertainties and thoughtful problem reformulation. Thus, the presence of large uncertainty is not a statement of poor model quality but an opportunity for gathering new data/knowledge in sparsely populated data spaces—paving the way for a paradigm shift in corrosion science and informatics.
Integrated computational material engineering to design new structural and functional materials with enhanced corrosion properties or to discover those with altogether new unforeseen properties will require additional advances in the above approaches. This challenge and opportunity is even more difficult in the case of compositionally complex alloys (CCAs),(3),31-32 corrosion inhibiting chemicals, coatings, and glasses33 where the possible combinations of elements or molecules are nearly limitless.
Consider a corrosion-resistant alloy (CRA) such as a CCA or multiple principle element alloy (MPEA)(4) that forms a passive film and is susceptible to passive film break down and local corrosion. Engineers might be interested in designing to reach target properties or predicting the properties for an existing alloy. A plethora of variables govern just this one mode of corrosion given the many degrees of freedom in composition. Design using trial-and-error approaches to explore many compositional variables is intractable and expert intuition is insufficient as well given the complex nature of the alloy and the corrosion phenomenon.
It is envisioned that big data and artificial intelligence in the form of ML approaches will provide insights and relationships that might be difficult to detect in order to rationally tailor alloy compositions to produce desired corrosion properties.
The need for ML approach is motivated by the fact that the MPEA chemical and composition search space is vast (>106 possibilities) and only a tiny fraction has been experimentally documented.34-35 Design rules that govern corrosion properties as a function of MPEA composition and microstructure are not known a priori. The desired properties are likely a function of key descriptors that reflect how the atomic and microstructure scale constituents impact the corrosion behavior.
A dataset of attributes and properties of corrosion control materials, coatings, as well as corrosion inhibitors can be built from surveying the literature based on a finite number of experiments. Such a dataset will need to be built to contain information on composition, processing history, structure-phase, and bulk environment used for testing the corrosion properties as well as the measured behavior or outcomes pertinent to pitting corrosion. Each MPEA composition, in turn, may be represented using phenomenological atomistic, microscale, and mesoscale descriptors. Other descriptors can be added as well, such as essential features that describe the protective oxide and the bulk, interface, and local environments. New MPEAs with promising targeted corrosion properties could be predicted and subsequently experimentally validated in this manner. Feedback could be used for iterative improvement of the ML models and refinement, as well as down-selection of the descriptors producing certain outcomes.36 However, it is non-trivial to discover “new physics” using such approaches. Nevertheless, ML might help identify critical relationships between key attributes and outcomes that could lead to directed further scientific inquiry that in turn does uncover new underlying physics.
The idea of designing materials and other means of corrosion control by ML and other such approaches provides a promising snapshot of one “new frontier” in corrosion control that might enable a surge in progress. As these approaches continue to be developed and utilized, we may discover new opportunities and challenges. Let’s continue to pursue these grand challenges and expand the frontiers together within CORROSION.
An assistant professor of materials science and engineering and mechanical engineering at UVa.
Figure 1 in the Frank Newman Speller Award paper by N. Sridhar4 presents a nice summary of the range of approaches and models useful for predictive analysis of corrosion.
One of the unique attributes of CCAs is that they contain multiple (at least five) principal alloying elements of nearly equi-atomic concentration and yet have a global crystal structure with well-defined Bragg reflections indicative of long-range order. Unlike the traditional metallic alloys that have complex microstructure (including formation of precipitates), HEAs are typically solid solutions of face-centered cubic (fcc), body-centered cubic (bcc), or hexagonally closed packed (HCP) phases. The microstructures can be tuned by heat treatments such that mixed phases (e.g., bcc+fcc) are also obtained.
MPEAs are a subset of CCAs.
John R. Scully was supported by the Center for Performance and Design of Nuclear Waste Forms and Containers, an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences under Award No. DE-SC0016584. Helpful discussions with James Saal at Citrine Inc., Wolfgang Windl at The Ohio State, and Pin Lu at Questek Corp. are greatly appreciated.