A framework toward knowledge transfer across heterogeneous and sparse datasets for fatigue life estimation of additively manufactured alloys

Despite rapid advancement and widespread applications of machine learning (ML) and artificial intelligence (AI) methods, interpreting sparse observations across heterogeneous and non-overlapping domains remains a fundamental challenge. This study proposes two methods to work in tandem to address these challenges in fatigue life estimation of additively manufactured alloys- (a) data imputation using the PRISM-MICE algorithm for sparse datasets, and (b) Separable Interpolating Transfer Learning (SeITL) which introduces - for the first time - a separable and scalable transfer learning approach for scientific computing. SeITL is inspired by the recently developed Interpolating Neural Network (INN), which incorporates variable separation, Galerkin approximation, and tensor decomposition into a graph data structure. The article discusses at length how to handle datasets collected from various sources, how to prudently screen the data and choose features to impute, imputation based on the statistical distribution of the existing dataset with numerical, categorical, and qualitative features, and transferring knowledge from and to non-overlapping feature spaces without resorting to any intermediate latent space. PRISM-MICE and SeITL achieve 2–7 times greater accuracy compared to multilayer perception (MLP) models on the imputed dataset. Moreover, the dataset is chosen in a way that, without imputation, it is not possible to consider all the features for knowledge discovery. The proposed work is not limited to the manufacturing/fatigue life estimation domain and can be extended without any major modifications to other scientific and engineering applications.