Denoising diffusion models for inverse design of inflatable structures with programmable deformations

Sara Karimi, Nikolaos N. Vlassis · 2025

Read Denoising diffusion models for inverse design of inflatable structures with programmable deformations by Sara Karimi, Nikolaos N. Vlassis (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-2508-13097.txt.

Introduction

Sara Karimi and Nikolaos Vlassis present a 2025 method for designing inflatable structures that take a wanted shape when pumped up. Soft robots, deployable space hardware, and medical devices all use pressure-driven bending. The hard part runs backward: given the final inflated shape, find the starting shape and material layout that will produce it under pressure.

Their framework treats that backward problem as conditional generation. A denoising diffusion model, the family behind modern image generators, receives descriptors of the target shape and outputs a picture-like map of the uninflated design. A fixed pipeline turns those maps into simulations and back. Tests with simple and richer descriptors produced many candidate designs quickly, each inflating into the target form. The authors highlight parallel exploration of designs under complex constraints.

Catalog record

CREATORSara Karimi, Nikolaos N. Vlassis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-13097

Browse the full library · About the Tower of Babel

Full text

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

Denoising diffusion models for inverse design of inflatable structures with programmable deformations Sara Karimi, Nikolaos N. Vlassis Abstract Programmable structures are systems whose undeformed geometries and material property distributions are deliberately designed to achieve prescribed deformed configurations under specific loading conditions. Inflatable structures are a prominent example, using internal pressurization to realize large, nonlinear deformations in applications ranging from soft robotics and deployable aerospace systems to biomedical devices and adaptive architecture. We present a generative design framework based on denoising diffusion probabilistic models (DDPMs) for the inverse design of elastic structures undergoing large, nonlinear deformations under pressure-driven actuation. The method formulates the inverse design as a conditional generation task, using geometric descriptors of target deformed states as inputs and outputting image-based representations of the undeformed configuration. Representing these configurations as simple images is achieved by establishing a pre- and postprocessing pipeline that involves a fixed image processing, simulation setup, and descriptor extraction methods. Numerical experiments with scalar and higher-dimensional descriptors show that the framework can quickly produce diverse undeformed configurations that achieve the desired deformations when inflated, enabling parallel exploration of viable design candidates while accommodating complex constraints. 1 Introduction Programming elastic structures to achieve large, target deformations is central to applications requiring controlled shape change ( Oliver et al. 2016 ; Zhou and Sheiko 2016 ) , adaptive behavior ( Fan et al. 2016 ; Walther 2020 ) , or mechanical functionality ( McCracken et al. 2020 ; Fang et al. 2022 ) . Here, ”programming” refers to the deliberate design of undeformed geometries and material property distributions that produce prescribed deformed configurations under specific loading conditions. Mechanical metamaterials exemplify this principle: their unit cell geometries are tailored to produce pattern transformations that yield negative Poisson’s ratio ( Ren et al. 2018 ) , tunable stiffness ( Bertoldi et al. 2017 ) , or controlled energy absorption ( Fu et al. 2019 ) . Deployable structures in aerospace applications ( Zirbel et al. 2013 ; Santo et al. 2014 ; Zhang et al. 2021 ) – such as foldable antennas, booms, or membrane-based enclosures – are similarly designed to transition from compact stowed states to precise operational geometries through elastic deformation. In stretchable electronics and wearable sensors ( Rogers et al. 2010 ; Homayounfar and Andrew 2020 ) , substrates must undergo significant deformation while maintaining electrical and mechanical integrity ( Kim et al. 2008 ) . Across these systems, the key challenge lies in mapping undeformed configurations to desired deformed states in a way that accounts for geometric and material nonlinearities, complex loading, and boundary constraints ( Zheng et al. 2023 ; Jiao et al. 2023 ) . A prominent subset of such systems includes pneumatically actuated or inflatable structures ( Ali et al. 2009 ; Fenci and Currie 2017 ; Xavier et al. 2022 ) , which use internal pressurization to drive deformation and mechanical response. In soft robotics, elastomeric actuators with embedded air chambers bend ( Shepherd et al. 2011 ) , elongate ( Polygerinos et al. 2015 ) , or twist ( Connolly et al. 2015 ) in response to applied pressure, enabling locomotion, grasping, or shape adaptation ( Rus and Tolley 2015 ) . The structure of these actuators guides their deformation ( Gorissen et al. 2020 ) . For example, they often incorporate anisotropic reinforcements – e.g., helically wound fibers – to constrain expansion and produce targeted motions ( Bishop-Moser and Kota 2015 ) . In biomedical applications, balloon-expandable stents ( Dumoulin and Cochelin 2000 ) , inflatable catheters ( Choi et al. 2014 ) , and occlusion devices ( Gianturco et al. 1975 ) deploy from compact states to functional geometries within the body, requiring precise shape change under physiological constraints ( Timmins et al. 2008 ) . Similarly, architectural systems, such as deployable shelters ( Moy et al. 2022 ) and ETFE cushions ( Hu et al. 2017 ) , rely on controlled inflation to form tension-stabilized surfaces or load-bearing arches. The common thread in these applications is the targeted use of pressure-driven actuation to realize specific deformed geometries to perform given tasks, often under non-uniform constraints and with strong coupling between material properties, geometry, and loading paths ( Skouras et al. 2014 ; Gorissen et al. 2017 ) . Conventional approaches to inverse design of deformable structures – such as topology optimization ( Bendsøe and Kikuchi 1988 ) or heuristic iterative schemes ( Xie and Steven 1993 ) – typically require repeated calls to finite element method (FEM) simulations ( Rozvany 2001 ; Liu and Tovar 2014 ) , making them computationally prohibitive for large-scale design spaces ( Mukherjee et al. 2021 ) . These methods often rely on simplified material models ( Andreassen et al. 2011 ) or gradient-based updates that can struggle with non-smooth or highly nonlinear responses, particularly under large deformations ( Bruns and Tortorelli 2001 ) . Moreover, most existing frameworks assume simpler boundary conditions and domains shapes or have to adopt advanced methods to handle complex configurations ( Rozvany 2012 ; Wang et al. 2003 ; Yoon 2011 ; Guo et al. 2013 ) , limiting their applicability in systems with complex actuation, such as pressure-driven inflatable structures. The final deformed configurations result from complex interactions among initial geometry, boundary constraints, and material properties, often under non-uniform pressure loads and contact conditions ( Caasenbrood et al. 2020 ; Panetta et al. 2021 ) , which can violate assumptions commonly made in conventional design pipelines. To address these limitations, data-driven inverse design frameworks have emerged as alternatives to classical computational methods ( Regenwetter et al. 2022 ; Koul 2024 ; Cerniauskas et al. 2024 ) . Many of these leverage latent space generative models, such as variational autoencoders (VAEs) ( Kingma and Welling 2013 ) and generative adversarial networks (GANs) ( Goodfellow et al. 2014 ) among others, to learn mappings between design parameters and structural responses from simulated or experimental datasets. VAEs offer stable training dynamics and structured latent spaces, making them appealing for inverse design tasks ( Kim et al. 2021 ; Yonekura and Suzuki 2021 ) . However, they often require strong regularization to enforce smoothness and disentanglement in the latent space ( Mathieu et al. 2019 ) , which can reduce the fidelity of reconstructed or generated samples and may struggle to capture sharp, high-frequency features ( Asperti and Trentin 2020 ) such as localized material discontinuities. GAN-based models, on the contrary, can generate sharper and more detailed outputs, making them attractive for applications where fine geometric fidelity is critical ( Yang et al. 2018 ; Mao et al. 2020 ) . However, they are prone to training instabilities, including mode collapse and sensitivity to hyperparameters arising from the need to maintain a delicate balance between generator and discriminator ( Saxena and Cao 2021 ) . Recent studies attempt to address these limitations by incorporating physics-informed loss functions and hybrid discriminator designs ( Kazemi et al. 2022 ; Nguyen et al. 2022 ) . In parallel, denoising diffusion probabilistic models (DDPMs) ( Ho et al. 2020 ) have been replacing state-of-the-art methods in synthesis tasks and, by extension, have gained traction in the inverse design of structures ( Vlassis and Sun 2023

Read the complete text