The performance degradation of materials exposed to corrosive atmospheric environments is a serious problem. Corrosion maintenance strategies and cycles are informed largely by historical trends in corrosivity dependent on the location of interest. Assessments are often made with sparse experimental data and are generalized to certain materials and baseline conditions observed in the past. The development of a predictive model which can better inform maintenance strategies and cycles offers opportunities for time savings and cost avoidance. Such a model would need to predict corrosion damage accumulation as a function of material of interest, location of interest, and atmospheric conditions during the period of exposure of interest.

Machine learning has recently gained attention as a way to model complex systems and processes. Atmospheric corrosion is certainly a complex process as there are many parameters that influence it which are uncontrolled and vary in difficult to predict ways. A machine learning algorithm would require data to learn from which might include environmental conditions such as temperature, relative humidity, UV light exposure, time of wetness, salt concentration, applied loads, and electrochemical data for substrates, inhibitors, and coatings. The ability of machine learning to determine models and relationships from large data sets is well demonstrated; however, a Bayesian model approach can be more suitable when data is limited, expensive or time consuming to obtain, and expert knowledge can be leveraged to construct the relationships. While certain meteorological variables are well categorized and easy to retrieve other variables like corrosion damage or salt deposition are sparsely reported if measured at all. For this reason, a Bayesian Network Model has been developed to better predict the corrosion damage accumulation of C1010 steel and three different aerospace coating systems applied over AA2024-T351.

The network map was constructed from parameters (nodes) which are known to influence corrosion. Mathematical relationships were used to develop relationships between nodes or to calculate new nodes as available. Finite element analysis was used to refine the network map and provide supplemental inputs for features that are difficult to measure experimentally. The model was initially trained from historic corrosion data gathered from collaborators across the DoD. Contemporary experimental data was also gathered in a 19-site survey of field exposure sites across the world. This dataset provided further inputs for training and also data for validation and testing. The methodology behind model construction will be highlighted and model predictions will be compared against experimental results. To date, there has been good correlation between model predictions and experimental results for C1010 steel mass loss, which provides a basis to further extend the model to aerospace coating systems.