PNNL

Abstract

We present a Knowledge-Informed Mapping toolkit in Python programming language, named KIM, to optimize the development of the mapping from a vector of inputs X to a vector of outputs Y. KIM builds on the methodology development of deep learning-based inverse mapping in Jiang et al. (2023) and A. Wang et al. (2025). It involves two key procedures: (1) an exploratory data analysis using information theory to identify the dependency between X and Y and filter out both insignificant and redundant inputs through global sensitivity analysis and conditional independence testing; and (2) ensemble learning of using neural networks to account for its structural uncertainty. KIM offers a preliminary understanding of data interdependencies while optimizing the training step with uncertainty accounted for. We demonstrate the applications of KIM through two use cases of developing inverse mappings for learning the parameters of an integrated hydrological model and a large eddy simulation cloud model. We expect this toolkit will be helpful to glue the model data integration for Earth science applications.