<p>The acceleration, velocity, and displacement response spectra are critical in the seismic design and dynamic analysis of structures. In many practical cases, only one type of spectrum is available to characterize the ground motion. To obtain the other required response spectra from the provided one, adopting relationships that map these spectra between one another offers a practical solution. However, although many models for such relationships have been proposed, their applicability and accuracy are limited due to the simplified regression techniques employed. To address these limitations, this study develops a set of models for estimating the relationships among these three response spectra based on machine learning that have been successfully applied across various fields. Specifically, the long short-term memory neural networks, one of the widely adopted algorithms in machine learning, are utilized to construct the predictive models, owing to their superior learning and information-selection capabilities. The Bayesian optimization algorithm is used for model hyper-parameters selection, as its effectiveness has been well-demonstrated even in complex optimization problems. A dataset of 16,660 horizontal seismic acceleration records is utilized, with magnitudes of 4.0–9.0 and epicentral distances of 10–200 km. The ground motion dataset was collected at 338 stations across four site classes. Finally, the proposed predictive models are systematically compared with existing models, and the results indicate that the proposed models provide improved accuracy and more stable performance.</p>