In recent years, various computational approaches such as Calculation of Phase Diagrams (CALPHAD), molecular dynamics simulation (MD), and density functional theory (DFT) calculations have been utilized to estimate phase and to predict structural and electronic properties. However, these approaches have limitations due to high computational costs and accuracies.
In the last decade, Machine Learning (ML) methods (specifically, Artificial Neural Networks or Deep Learning, DL) have proven to be well suited to handle large data sets with large numbers of input parameter space in order to capture underlying features of a problem at hand. DL models have been successfully applied in areas such as image recognition, natural language processing, self-driving cars, anomaly detection, predictive modelling, among others. In recent years, ML has been applied in the field of CCAs as a data-driven feature learning approach. Several recent works have shown promising results for estimating physical properties of CCAs. For example, Wen et al. proposed a strategy to search for new CCAs with higher hardness values in the AlCoCrCuFeNi alloys by training ML models with measured physical properties of known CCAs. Other works suggested several Deep Learning models to predict phases of CCAs, where the existing empirical phase estimation rules can be modelled by ML/DL models.
Among the technical properties of interest, the coefficient of thermal expansion (CTE) is regarded as the most relevant property as the CTE of the coating and the substrate must not differ by more than 2 ppm. In general, the CTE of a material depends on the bonding strength and thus on the respective neighbouring atoms/ions (radius, valence, etc.) and their relative position described by the crystal structure. It can vary with direction for anisotropic materials and is temperature-dependent. To date, it is not possible to accurately calculate the CTE of a given material ab initio. However, use of Machine Learning models can help get good estimates. The application of ML to predict material combinations with a target CTE would be very beneficial in the development of a suitable coating combination.
There are several key obstacles with ML-guided CCA/CCC property predictions. In most cases, the key issue is the lack of good data. This issue can sometimes be slightly mitigated using regularization; however, having enough training data is essential to obtain a reliable model. The FORGE project partners will produce a large volume of experimentally measured datasets on CCA properties (Corrosion resistance, H2-embrittlement, Wear/Hardness) and CCC properties (Coefficient of thermal Expansion (CTE), Corrosion Rate) from a selected element pool. This will allow the development of ML/DL models that learn general features of the CCA/CCC compositions and their properties.
Another key issue is that DL models are inherently black-box models. It is often not clear why a model produces a specific result. When a DL model is estimating physical properties, it is important to have an explanation of the estimations of DL models. Explainable AI (XAI) can bring a lot of value in this regard. Explainable AI is a group of methods and approaches to explain results of complex machine learning models from the perspective of input features and output values.
There are various ways, namely statistical analysis, feature visualization, analysis of DL model weights, and counterfactual explanations, to explain DL models. It is important that the explanation methods are model-agnostic, where the methods work by analysing feature input and output pairs rather than internal details of the DL models.
One of the major goals of FORGE is to develop an Explainable AI based decision support system to help material scientists and engineers discover novel CCAs and CCCs with targeted properties with a minimum number of trial and errors.