Generative AI has become a powerful tool with growing relevance across industrial applications, offering the ability to explore novel and non-standard designs in real time. Among its foundational methods, autoencoders (AEs) are widely used for learning compact latent representations of complex data. However, traditional AEs often struggle with capturing smooth and interpretable latent spaces in challenging scenarios, resulting in unrealistic or non-physical outputs when interpolating. To address this, recent advances such as Variational Autoencoders (VAEs) [1] and Generative Adversarial Networks (GANs) [2] have emerged, though they still face limitations in ensuring consistency, smoothness, and physical interpretability of the latent space. In this work, we propose to use the Rank Reduction AutoEncoder (RRAE) [3], which enforces a low-rank constraint on the latent space via truncated SVD during training. This constraint promotes continuous and linearly interpolable latent spaces, enhancing the generative capabilities of the AE. We apply this framework to the challenge of handling components operating under high temperatures and designing effective cooling strategies based on their thermal state. Both the 2D geometry of the part and its associated heat distribution (provided by a numerical solver) are encoded, learning a correlation between them in the latent space. As a result, it becomes possible to generate new geometries and instantly obtain their corresponding temperature fields without running any additional simulations. This approach enables fast design exploration and decision-making in thermally critical applications, where the ability to predict and manage heat flow is essential.