<p>Accurate thermal error modeling is fundamental for machine tool thermal error compensation. However, traditional data-driven thermal error modeling methods rely solely on the experimental temperature and thermal error data, neglecting the underlying thermal structural mechanism. To incorporate both mechanism-driven and data-driven characteristics, we propose a novel reduced order model (ROM)-augmented neural network (NN) thermal error modeling method. Firstly, a rapid and precise ROM is developed to supplant the original, expensive transient thermal structural simulation. The ROM preserves the thermal structural mechanism through singular value decomposition (SVD) on a series of simulations. Given measured key temperatures, it can reconstruct the thermal displacement field in real time through truncated eigenmodes superposition leveraging predicted modal coefficients. Subsequently, introducing the output of the ROM as an additional pivotal variable for the middle layers of the NN, the ROM-augmented NN model can be trained with experimental data. Key temperature points, the input in thermal error prediction, are selected by combining <i>K</i>-means clustering with information gain. Finally, the proposed model demonstrates a 71%, 41%, 34%, and 21% relative improvement in performance across five spindle speeds compared to the ROM, multiple linear regression (MLR), least square support vector machine (LSSVM) and NN models, respectively.</p>