 Abstract machine learning force fields, MLFFs, are being developed to enable predictive simulations of realistic molecules at a fraction of the computational cost compared to AB Innocio methods. However, two key challenges remain, developing efficient descriptors for non-local interatomic interactions and reducing the dimensionality of these descriptors to increase their applicability and interpretability. In this paper, the authors propose an automated approach to substantially reduce the number of interatomic descriptors while maintaining accuracy and efficiency. They demonstrate this approach on the example of global GDML MLFFs, finding that non-local features are critical to retaining accuracy for peptids, DNA-based pairs, fatty acids, and supramolecular complexes. The number of required non-local features in the reduced descriptors becomes comparable to the number of local interatomic features, those within 5a, suggesting that global MLFFs can now be constructed whose cost increases linearly rather than exponentially with system size. This article was authored by Adil Kabilder, Valentin Vassilov-Golindo, Stefank Mieller, and others. We are article.tv, links in the description below.