A freeform imaging spectrometer design method is proposed. Only system specifications, including the slit length, numerical aperture, magnification, spectral range, and spectral resolution, are required as design inputs, and imaging spectrometer systems with various optical power distributions and structures are output. Using three-mirror imaging spectrometers as design examples, a series of initial solutions are obtained with a diffraction grating placed at the secondary mirror. In order to understand and explore the utilization of large number of output results, these initial solutions are visualized by mapping from high-dimensional solution space to three-dimensional space using the t-distributed symmetric neighbor embedding (t-SNE) dimensionality reduction algorithm; then a neural network is trained to fit the imaging qualities, showing good generalization performance, and can predict imaging quality for systems with unknown optical power distributions and structures.