Abstract
Ever since Yau's non-constructive existence proof of Ricci-flat metrics onCalabi-Yau manifolds, finding their explicit construction remains a majorobstacle to development of both string theory and algebraic geometry. Recentcomputational approaches employ machine learning to create novel neuralrepresentations for approximating these metrics, offering high accuracy butlimited interpretability. In this paper, we analyse machine learningapproximations to flat metrics of Fermat Calabi-Yau n-folds and some of theirone-parameter deformations in three dimensions in order to discover their newproperties. We formalise cases in which the flat metric has more symmetriesthan the underlying manifold, and prove that these symmetries imply that theflat metric admits a surprisingly compact representation for certain choices ofcomplex structure moduli. We show that such symmetries uniquely determine theflat metric on certain loci, for which we present an analytic form. We alsoincorporate our theoretical results into neural networks to achievestate-of-the-art reductions in Ricci curvature for multiple Calabi-Yaumanifolds. We conclude by distilling the ML models to obtain for the first timeclosed form expressions for Kahler metrics with near-zero scalar curvature.