Variational Rank Reduction Autoencoders for Generative Thermal Design

Alicia Tierz, Jad Mounayer, Beatriz Moya, Francisco Chinesta · 2025

Read Variational Rank Reduction Autoencoders for Generative Thermal Design by Alicia Tierz, Jad Mounayer, Beatriz Moya, Francisco Chinesta (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-08515.txt.

Introduction

Alicia Tierz and 3 co-authors build a design tool for heat flow that respects physics in a 2025 engineering paper. Good thermal design needs many trial shapes. Each trial costs a slow, costly simulation. Common generative models make that search worse. Their latent spaces are tangled and jumpy, so nearby points can give shapes that break the physics.

The team's hybrid pairs two networks. One keeps the latent space smooth and low-rank as it learns to compress shapes. The other is a deep operator network that rates heat behaviour fast. Designers could then search heatsinks and exchangers by browsing, not by waiting. A shape that looks fine but breaks heat flow is not a design. The frame is built to keep that line clear.

Catalog record

CREATORAlicia Tierz, Jad Mounayer, Beatriz Moya, Francisco Chinesta
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-08515

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Variational Rank Reduction Autoencoders for Generative Thermal Design Alicia Tierz, Jad Mounayer, Beatriz Moya, Francisco Chinesta Abstract Generative thermal design for complex geometries is fundamental in many areas of engineering, yet it faces two main challenges: the high computational cost of high-fidelity simulations and the limitations of conventional generative models. Approaches such as autoencoders (AEs) and variational autoencoders (VAEs) often produce unstructured latent spaces with discontinuities, which restricts their capacity to explore designs and generate physically consistent solutions. To address these limitations, we propose a hybrid framework that combines Variational Rank-Reduction Autoencoders (VRRAEs) with Deep Operator Networks (DeepONets). The VRRAE introduces a truncated SVD within the latent space, leading to continuous, interpretable, and well-structured representations that mitigate posterior collapse and improve geometric reconstruction. The DeepONet then exploits this compact latent encoding in its branch network, together with spatial coordinates in the trunk network, to predict temperature gradients efficiently and accurately. This hybrid approach not only enhances the quality of generated geometries and the accuracy of gradient prediction, but also provides a substantial advantage in inference efficiency compared to traditional numerical solvers. Overall, the study underscores the importance of structured latent representations for operator learning and highlights the potential of combining generative models and operator networks in thermal design and broader engineering applications. 1 Introduction Fast solvers and surrogates are essential for resolving temperature fields over complex geometries in different fields of engineering and scales. These applications range from mechanical elements and their manufacturing process chen2022data ; sajadi2024real ; pan2024temperature to urban analysis huo2022study ; kim2022estimating ; tanoori2024machine to ensure the performance, durability, and safety of the system. However, high-fidelity thermal simulations in practical applications, such as telecommunications and manufacturing, are often computationally expensive. Traditional solvers predominantly employ finite-element or finite-volume methods to predict thermal behaviour song20003 ; CHAQUET2025104385 . These methods are computationally intensive and time-consuming, especially when implemented in large-scale systems, real-time applications, or iterative design processes. The use of deep learning solvers results also inefficient since the design space often comprises thousands or more parameters zhou2024evolutionary ; Shan2010 . Consequently, reducing the dimensionality of the design space is essential for both simulation and generative optimisation. In this case, a major challenge in using deep learning solvers with reduction techniques is to effectively parameterize the design domain in the new manifold kashefi2022physics ; HE2024117130 . The development of surrogate models grounded in reduced-order models and machine learning, particularly deep learning, has been extensively investigated in recent years su2021super ; zhang2025study . Having surmounted the initial exploratory stages in black-box models, numerous alternatives in physics-informed machine learning are now available chen2025physics ; toscano2024inferring . In this context, deep operator networks (DeepONets Lu_2021 ) have emerged as an innovative and promising architecture for learning nonlinear operators directly from data. Their theoretical foundation is based on the Universal Approximation Theorem for Operators, which states that a neural network, even with a single hidden layer, can accurately approximate any continuous nonlinear operator. However, these solvers have not yet been comprehensively integrated with generative tools and spaces to thoroughly investigate solutions in optimisation and inverse problems. Generative artificial intelligence methodologies, particularly autoencoders, variational autoencoders, and generative adversarial neural networks, are employed to learn spaces, or manifolds. These methodologies facilitate the embedding of high-dimensional design spaces derived from high-dimensional parameter spaces, or even from images or NURBS, with the objective of determining a reduced-order parameterization kang2024physically ; wang2025generative . This reduced parameterization serves as a foundational tool for the exploration and analysis of the design space, including inverse analysis brzin2024using ; kang2024physically . Nevertheless, these methodologies, rooted in deep neural networks, notwithstanding their preliminary reconstruction metrics, yield discontinuous solution spaces. This phenomenon, often referred to as “holes”, pertains to regions within the manifold that, due to various factors, primarily data scarcity, fail to be reconstructed with precision and coherence. Consequently, the outputs within these regions are devoid of significance and fail to accurately represent the underlying design space bengio2013representation . Furthermore, traditional Variational Autoencoders (VAEs), although they address continuity through probabilistic modelling, can suffer from posterior collapse and tend to generate blurred images because KL divergence is their only form of regularisation. This study presents a thermal solver for shapes inspired by manufacturing and electronic systems. The solution space is developed upon a linearly interpolable manifold, which serves as an efficient generative tool seamlessly integrated with the simulation process. Analyzing these temperature distributions enables the optimization of device design to mitigate excessive heat, which can otherwise lead to performance degradation, or even mechanical failure. The foundation of this approach lies in the use of rank reduction autoencoders (RRAEs) mounayer2025rankreductionautoencoders , with a particular emphasis on their variational approximation (VRRAEs) mounayer2025variational . This methodology incorporates singular value decomposition (SVD) in conjunction with traditional autoencoders to enhance the quality of the reduced manifolds. As a result, the new reduced parametrization can be coupled with solvers, in this case DeepONets, to perform physics prediction of the designs from the reduced exploration space. Section 2 expands the state-of-art and challenges of generative AI and surrogate modeling. Section 3 states the methodology with a detailed description of the rank reduction autoencoder used and its combination with surrogates to predict the temperature fields of the geometries. Results are presented in section 4, demostrating the outperformance of the proposed method with respect to other state-of-art strategies. To finalize, we present the conclusions and future perspectives of this work for the further development and use of the presented approach. 2 Related Work Generative Artificial Intelligence Generative Artificial Intelligence (Gen-AI) constitutes a significant challenge within the current landscape of artificial intelligence research, and particularly in engineering applications regenwetter2022deep . It can be argued that the foundation of these methodologies is rooted in the techniques associated with reduced order modeling, from POD and non-linear manifold learning to autoencoders (AE) loaiza2024deep . In essence, these techniques learn from data the spaces where the problem object of study is embedded, thereby facilitating the exploration of the manifold to identify further potential solutions regarding a new objective. Autoencoders are among the most utilized tools due to their superior performance in the presence of nonlinear correlations within data. However, vanilla autoencoders struggle in learning effective representations in a complex setting, which can often result in a latent space suffering from “holes”, i.e. areas where the corresponding manifold is undefined, and its prediction does not produce physically meaningful predictions bengio2013representation . There exist two main methodologies that attempt to address this challenge. Variational Autoencoders (VAEs) kingma2013auto extend the concept of traditional autoencoders by applying Bayes’ theory. With this constraint, the AE does not learn a deterministic space, but a distribution that provides continuity and smoothness to the manifold of solutions. This technique has been applied in a variety of fields. The authors in ( wang2025generative ) exploit the capabilities of VAEs for aerofoil design from NURB entries. Similarly, Guo et al. ( guo2018indirect ) apply this technique to topology optimisation for heat conduction design from images. Yang et al. ( yang2024research ) use \beta -VAEs to enhance the interpretability of the manifold in multi heat source designs of electronic devices, applying a parametrization of the domain to perform forward analysis. Generally, the efficacy of VAEs is heavily reliant on the

Read the complete text