4D-printed devices with shape memory alloy (SMA) components are promising for a wide range of applications due to their programmability, high energy density, and lightweight nature [1]. However, accurate prediction of their functional behaviour remains challenging because of the complex, thermally dependent phase transformation of SMAs [2]. Classical finite element analysis (FEA) can accurately model the constitutive behaviour of SMAs and the mechanical deformation of 4D-printed devices. Nevertheless, its low computational efficiency and reliance on trial-and-error for convergence constitute a major research gap, limiting its applicability to computationally intensive tasks such as system-level design, simulation, and optimization. This paper presents a physics-informed neural network (PINN)–based framework for the efficient simulation of 4D-printed devices. By embedding governing equations, phase-transformation evolution constraints, and boundary conditions into the training loss, the underlying SMA physics is consistently preserved in the PINN model. The proposed framework is further applied to structural topology optimization, where the PINN provides robust predictions of the thermoelastic response while avoiding the mesh distortion issues that commonly arise in FEA-based optimization. Moreover, the model is inherently differentiable with respect to the design parameters, greatly simplifying sensitivity computation. The effectiveness of the proposed method is validated through numerical benchmarks under isobaric and isothermal loading conditions. The results demonstrate that the method maintains both robustness and predictive accuracy even in large-deformation regimes. The PINN-based surrogate model thus offers a promising tool for the structural optimization of 4D-printed SMA devices.