Do Diffusion Models Dream of Electric Planes? Discrete and Continuous Simulation-Based Inference for Aircraft Design

Aurelien Ghiglino; Daniel Elenius; Anirban Roy; Ramneet Kaur; Manoj Acharya; Colin Samplawski; Brian Matejek; Susmit Jha

cs.LGFeb 27, 2026

Do Diffusion Models Dream of Electric Planes? Discrete and Continuous Simulation-Based Inference for Aircraft Design

Authors: Aurelien Ghiglino, Daniel Elenius, Anirban Roy, Ramneet Kaur, Manoj Acharya, Colin Samplawski, Brian Matejek, Susmit Jha,

Organizations: Computer Science Laboratory, SRI · Aerospace Design Laboratory, Stanford University

Abstract

In this paper, we generate conceptual engineering designs of electric vertical take-off and landing (eVTOL) aircraft. We follow the paradigm of simulation-based inference (SBI), whereby we look to learn a posterior distribution over the full eVTOL design space. To learn this distribution, we sample over discrete aircraft configurations (topologies) and their corresponding set of continuous parameters. Therefore, we introduce a hierarchical probabilistic model consisting of two diffusion models. The first model leverages recent work on Riemannian Diffusion Language Modeling (RDLM) and Unified World Models (UWMs) to enable us to sample topologies from a discrete and continuous space. For the second model we introduce a masked diffusion approach to sample the corresponding parameters conditioned on the topology. Our approach rediscovers known trends and governing physical laws in aircraft design, while significantly accelerating design generation.

Figures & tables

Figure 1: Summary of the design tree and design parameters, refer to App. E , Fig. 19 for a more detailed image.

Figure 2: MixeDiT-MaskeDiT architecture. Models are summarized on the left, with the full design pipeline highlighted on the right.

Figure 3: Comparison of the marginal test distribution of the topologies vs. the marginal samples generated by MixeDiT. The log scale highlights the performance on low-probability topologies.

Figure 4: Summary statistics of generated samples using MixeDiT. The original data distribution (blue) closely matches the MixeDiT marginal generated samples (orange). Conditioning on higher mass designs (green,

+2\sigma_{mass}

) significantly increased the number of two-winged designs, while conditioning on a low mass design dramatically reduces the number of two-winged designs (red,

-2\sigma_{mass}

Metric	MaskeDiT	NLE
MMD Value	$0.004\pm 0.001$	$0.002\pm 0.001$
Joint C2ST	$0.777\pm 0.040$	$0.744\pm 0.023$
Mean Marg. C2ST	$0.614\pm 0.016$	$0.596\pm 0.010$
Max Marg. C2ST	$0.770\pm 0.076$	$0.818\pm 0.043$

Table 1: Likelihood comparison between generated samples

p_{\phi}(\mathbf{x}|\mathbf{\bm{\theta)}}

and test data

p_{\mathcal{D}}(\mathbf{x}|\mathbf{\bm{\theta)}}

. Average is taken over

10

topologies with more than

200

samples in the test dataset.

Panel of Figure 5 — Figure 5: Posterior predictive distributions of selected $\mathbf{x}$ variables under three conditioning scenarios.

Figure 6: Case Study A: Trends in selected variables for increasing total eVTOL mass with fixed topology, drag and state of charge.

Figure 7: Case Study B: Trends in selected variables for biplane (Bi) and monoplane (Mono) eVTOLs for increasing lift ( $C_{L}$ ) while keeping wing weight fixed.

Figure 8: Case Study B: parameter variations when considering biplane and monoplane topologies.

Figure 9: Case Study C: Trends in selected variables for increasing eVTOL total mass while keeping other observations such as lift, drag, wing mass and state of charge fixed.

Figure 10: Case Study D: Trends in selected variables for increasing drag while keeping the remaining observations (e.g. total mass, lift, state of charge) fixed.

Figure 11: Case Study D: rendered topology changes while conditioning on a range of drag coefficients.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

M_{C}\sim\tfrac{1}{5}\,\text{Bernoulli}(0.3)+\tfrac{1}{5}\,\text{Bernoulli}(0.7)+\tfrac{1}{5}\,\delta_{\mathbf{1}_{d_{\theta}+d_{x}}}+\tfrac{1}{5}\,\delta_{[\,\mathbf{0}_{d_{\theta}},\,\mathbf{1}_{d_{x}}\,]}+\tfrac{1}{5}\,\delta_{[\,\mathbf{1}_{d_{\theta}},\,\mathbf{0}_{d_{x}}\,]},

Appendix

Algorithm 3 MaskeDiT Training

t_{k}\leftarrow 1-\frac{k}{K}(1-T_{\min})

Appendix

Algorithm 4 MaskeDiT sampling process

Figure 12: Case Study A: The pairplot for several selected variables for increasing mass.

Figure 13: Case Study A: Note the correlation between seeds. This collage highlights the diversity of available design choices, whereby there are significant differences in battery mass across these designs despite similar CAD representations.

Figure 14: Case Study B: The pairplot for several selected variables comparing monoplanes and biplanes with increasing lift.

Panel of Figure 16 — Figure 16: Case Study C: Several selected variables for increasing mass without fixing topology.

Figure 17: Case Study C: renderings for increasing mass and allowing variable topology.

Panel of Figure 18 — Figure 18: Case Study D: Several selected variables for increasing drag without fixing topology.

Figure 19: Full design tree for an example aircraft.

Component	Description	Distribution
RotorTemplate	Propeller data class.
tip_radius		$\mathcal{U}(0.5,1.5)$
hub_radius		$\mathcal{U}(0.1,0.2)$
number_of_blades		$\mathcal{C}(p),p=0.25$ classes: (2,5)
design_tip_mach		$\mathcal{U}(0.5,0.8)$
design_altitude		$\mathcal{U}(500.0,2000.0)$

Appendix

Table 2: Description of probabilistic generator (prior) over aircraft model.

Figure 20: The full posterior predictive distribution over all ten $\mathbf{x}=\bm{\mu}$ variables. The average aircraft in the dataset, shown with the solid red line, is what the sampling from the architecture is conditioned on. When the samples are run through the simulator again, the distributions of $x$ values obtained are shown in blue, with the dotted black line showing the means of those distributions. Data distributions are in gray.

Figure 21: The full posterior predictive distribution over all ten $\mathbf{x}=\bm{\mu}+\bm{\sigma}$ variables.

Figure 22: The full posterior predictive distribution over all ten $\mathbf{x}=\bm{\mu}-\bm{\sigma}$ variables.

Metric	MaskeDiT	NLE
MMD Value	$0.0040\pm 0.0006$	$0.0021\pm 0.0007$
Joint C2ST	$0.7773\pm 0.0395$	$0.7445\pm 0.0227$
Mean Marginal C2ST	$0.6142\pm 0.0155$	$0.5961\pm 0.0105$
Median Marginal C2ST	$0.5973\pm 0.0299$	$0.5614\pm 0.0158$
Max Marginal C2ST	$0.7699\pm 0.0762$	$0.8185\pm 0.0426$
Min Marginal C2ST	$0.5065\pm 0.0247$	$0.4963\pm 0.0210$

Appendix

Table 3: Likelihood comparison between generated samples

p_{\phi}(\mathbf{x}|\mathbf{\bm{\theta)}}

and test data

p_{\mathcal{D}}(\mathbf{x}|\mathbf{\bm{\theta)}}

. Average is taken over

10

topologies with more than

200

samples in the test dataset, Mean and standard deviation

\mu\pm\sigma

shown.

Diffusion Models

Diffusion models learn to reverse gradual data corruption, enabling flexible generation, editing, forecasting, and inference.

See all 1,169 papers

AlphaJet: Automated Conceptual Aircraft Synthesis via Disentangled Generative Priors and Topology-Preserving Evolutionary Search
Apr 29, 2026Boris KriukVariational Autoencoders 3D Shape Generation
Hong Kong University of Science and Technology Clear Water Bay, Hong Kong, HKSAR
Diffusion Models in Simulation-Based Inference: A Tutorial Review
Dec 22, 2025Jonas Arruda, Niels Bracher, Ullrich Köthe +2Simulation-Based Inference Parameter Estimation
Bonn Center for Mathematical Life Sciences, Life & Medical Sciences Institute, University of Bonn, Germany · Center for Modeling, Simulation, & Imaging in Medicine (CeMSIM), Rensselaer Polytechnic Institute, NY, USA · Computer Vision and Learning Lab, Heidelberg University, Germany
Zero-shot rib design: merging training-free generative prior with topology optimization
Sep 9, 2026Yongmin Kwon, Namwoo KangTopology Optimization Structural Mechanics
Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, 193 Munji-ro, Yuseong-gu, Daejeon, 34051, Republic of Korea · AI Research Team, Narnia Labs, 193 Munji-ro, Yuseong-gu, Daejeon, 34051, Republic of Korea

Apr 29, 2026J/K move · Enter open · S save

AlphaJet: Automated Conceptual Aircraft Synthesis via Disentangled Generative Priors and Topology-Preserving Evolutionary Search

Boris Kriuk

Hong Kong University of Science and Technology Clear Water Bay, Hong Kong, HKSAR

Variational Autoencoders 3D Shape Generation Conditional Generative Modeling Genetic Algorithms

Open paper AskPDF arXiv

Conceptual aircraft design is traditionally an expert-mediated iterative process in which a human designer proposes a configuration, runs low-order physics, inspects the result, and re-proposes. We present AlphaJet, an end-to-end automated synthesis pipeline that closes this loop. From a textual mission specification (mass, range, cruise speed, hard size envelope, engine count, areal density) AlphaJet evolves a feasible 3D aircraft in real time, scored by a transparent multi-disciplinary fitness function covering aerodynamics, structures, weights, stability, packaging, and geometric mount consistency. Three contributions distinguish our approach: (i) an Anatomically-Disentangled Variational Autoencoder (AD-VAE) whose first 25 latent dimensions are supervised to align with named anatomical parameters, providing an interpretable shape prior; (ii) a topology-elitist genetic algorithm that protects the best individual from each of five tail topologies and triggers stagnation restarts, preventing premature collapse to a single configuration; and (iii) mount-aware geometric scoring that computes signed penetration between engines and other structural parts, eliminating the redundant artifacts common in generative aircraft models. The full loop runs interactively on a CPU and streams every generation to a browser viewer, making it a practical real-world automation tool for early-phase design-space exploration.

Abstract

Diffusion Models

Explore similar work

AlphaJet: Automated Conceptual Aircraft Synthesis via Disentangled Generative Priors and Topology-Preserving Evolutionary Search

Diffusion Models in Simulation-Based Inference: A Tutorial Review

Zero-shot rib design: merging training-free generative prior with topology optimization