Data-driven prediction of atomic-oxygen erosion yield in polymers from bulk physical properties

Abstract

Atomic oxygen (AO) erodes polymer surfaces in low Earth orbit, but durability data are limited. We evaluated a regression ensemble trained on 37 MISSE-2 polymers using eight selected bulk properties and five elemental-composition ratios. Its archived leave-one-out R^2(log) was 0.768; this estimate is conditional on feature selection using the complete training table. For nine post-MISSE-2 polymers, the ensemble gave R^2(log) = -0.037 and all predictions were within a factor of two. The training arithmetic-mean baseline also met that tolerance for all nine, highlighting the limited discrimination provided by their narrow target range. Ground-based comparisons included four PLAVIS grades. For an out-of-training mystery polymer with undisclosed composition, the predicted erosion yield was within a factor of 1.92 of the ground-based measurement. Conflicting PLAVIS measurement records prevent a definitive residue-erosion ordering or ground-validation score. Thermal residue, SEM contrast and surface chemistry provide distinct observational evidence; they do not establish a protective-layer mechanism. The results support further evaluation of orbital-data screening models with reconciled measurement records and prospectively fixed inputs..