Nonlinear response-history analysis (NRHA) is widely recognized as the most comprehensive method for representing dynamic behavior under earthquake. However, it leads to considerable computational cost, making it difficult for engineers to implement in practice, let alone use it as a foundation for exploring preliminary structural designs. Bridging the gap between the advantages of NRHA and its high computational overhead has long been a primary focus of academic research. To overcome the limitations of NRHA, this study uses graph neural networks (GNNs) to deconstruct structural systems. These are coupled with the long short-term memory (LSTM) networks to approximate static and nonlinear dynamic responses, specifically capturing time-delay effects. The data-driven approach genuinely learns the component configuration and force transmission of a structure, enabling the model to generalize the mechanical characteristics of structural systems while significantly reducing the computational cost of nonlinear analysis. Based on a fast and scalable machine-learning-based structural analysis model, this study further explores how such models can be integrated into the structural design decision-making process. Reinforcement Learning (RL) and the Nondominated Sorting Genetic Algorithm (NSGA) are investigated for constructing single- and multi-objective optimization frameworks. Through this integration, NRHA can fully realize its potential - not merely serving as a final-stage verification tool in the design process, but instead actively guiding design decisions throughout the structural design workflow, with the potential to help engineers better understand structural behavior and accomplish their tasks more effectively. The systematic incorporation of machine learning techniques is expected to contribute to the development of a new generation of structural analysis and design paradigms and workflows.