A general framework for extrapolation-aware prediction reliability in forward and inverse analyses of Gaussian mixture regression models

Abstract

Gaussian mixture regression (GMR) and direct inverse analysis (DIA) are powerful tools for molecular, material, and process design, but their reliability decreases when extrapolation occurs beyond the training data. To address this challenge, an index of extrapolation (IoE) that evaluates extrapolation potential and prediction trustworthiness in both forward and inverse analyses of GMR models is proposed. Defined as the negative logarithm of the probability density function, the IoE enables stable identification of interpolation-like and extrapolation-like regions. Validation with numerical simulations confirmed its ability to distinguish these regions, while applications to data sets of organic solubility, inorganic superconductivity, and batch processes demonstrated that high-IoE regions correspond to reduced prediction reliability, where prediction errors may show larger dispersion. The proposed IoE provides a practical and generalizable framework to assess applicability domains, enhancing the interpretability and reliability of machine learning models. This advancement strengthens data-driven design and supports more efficient discovery of novel molecules, materials, and processes.

Graphical abstract