ABSTRACT A novel encoding strategy, using molecular weight (MW) to represent the type of sintering additives, was proposed to perform predictions of thermal conductivity of porous silicon carbide ceramics by machine learning in the consideration of including the scientific meaning. For comparison, a conventional strategy, labeling those parameters with random numbers, was used. The predictive model built with MW encoding produces similar prediction accuracy with the one built with label encoding when the test data is randomly split. Permutation importance ranked total porosity as the dominant factor when using both encoding strategies. In addition, process‐only model and two‐stage model were proposed to support process design. With the presence of porosity in the explanatory variables, the two‐stage model showed improvements in error rates. The principal component analysis results further show that porosity reduced the coverage of the first two components, indicating an additional, partly independent source of variation. Leave‐one‐study‐out validation across seven literature groups of the primary model reveals large aggregate errors were driven by a single ultralow and narrow‐range cohort. Errors markedly improved either by excluding this cohort or by using a log‐scale target while retaining it. This framework combines novel encoding along with realistic cross‐study generalization for process‐aware modeling.
Chung et al. (Fri,) studied this question.