Stephen Brown¹, Marjorie Cavarroc², Manuel Mendez³, Ludvik Martinu¹, Jolanta E. Klemberg-Sapieha^1
1Polytechnique Montréal, Montréal, Québec, Canada
²Safran Tech, Magny-les-Hameaux, France
³MDS Coating Technologies, Montreal, Quebec, Canada
Solid particle erosion (SPE) is a phenomenon in which material is removed from a surface by a stream of impinging erodent particles entrained in gas. It presents a major issue in aerospace, particularly for the leading edges of compressor blades exposed to sand, dust, and ash particles. Though SPE has been studied to some degree for over 100 years, there remains a lack of understanding of how different conditions and material properties affect erosion. A review of the literature reveals this gap: while the existing body of work is useful for qualitative comparisons, inconsistencies in testing parameters across studies make quantitative comparisons challenging, if not impossible. To address these challenges, our research leverages machine learning (ML) algorithms, which are well-suited for handling the inherent variability in specific values. ML algorithms can estimate the influence of parameters even in the presence of inconsistent testing conditions, provided that sufficient high-quality data is available.
The present study compiles a database of over 1000 test results from the literature covering the SPE of metals and uses this to assess erosion predictability and parameter influence. Several algorithms are tested, including nearest neighbors, tree-based methods, multivariable regression, and neural networks. More than 40 features are used, including a full range of mechanical properties for particle and target materials, as well as oft-ignored system properties (e.g., working distance, nozzle properties) and article metadata (year of publication, compliance with standards). In cases where model interpretability is a problem, the use of Local Interpretable Model-Agnostic Explanations (LIME) is explored. The results quantify both the ability and limits of ML in predicting SPE under diverse experimental conditions and highlight the importance of incorporating both material properties and system features for accurate estimation of erosion.