Generative Design with Real-time Physics Explainability: Towards Product Definition in the Age of AI
Authors: M. de Campos, C. Drygala, E. Ross, A. Bruce, M. Bohon, H. Gottschalk
Organizations: Institute of Mathematics, TU Berlin, Straße des 17. Juni 135, 10623 Berlin, Germany · Hermann Föttinger Institute, TU Berlin, Straße des 17. Juni 135, 10623 Berlin, Germany
In this article, we explore opportunities from using generative AI in mechanical engineering. On the one hand, generative methods based on invertible neural networks have been recently proposed for the solution of inverse design problems, allowing to infer design parameters from performance labels. On the other hand, generative AI models have shown that they can act as real time capable surrogates for physics based simulations. Here we suggest that these two streams should be combined in a unified framework, where INN are used to control the design process from the customer requirements perspective and the generative simulation of the systems physics adds a layer of explainability, why the proposed design fulfils the requirements. As generative design is prone to hallucinations, we propose uncertainty quantification to fend off physically infeasible requests. Finally, validation with classical tools is a requirement for generative design process. To demonstrate the capabilities of our novel generative design paradigm, we develop two interactive environments, one for airfoil design based on simulator data and one for the combustor design space exploration based on measured data.
Figures & tables
Figure 1: Generative design with real-time physics explainability. Lift and drag targets condition Diagonal Flow Matching (Diag-CFM) to generate an airfoil; a learned surrogate predicts the physical field. Geometry, physical behaviour, and uncertainty are inspected together. All visual elements are schematic.
Turbulence model
Airfoil class
Airfoils
Mach
Re ( ×106 )
AoA ( ∘ )
Simulations
SA (fully turbulent)
FT
30,000
[0.1,0.85]
[1,10]
[−2,6]
400,000
SA (fully turbulent)
NLF
4,800
[0.1,0.85]
[1,10]
[−2,6]
50,000
SA + eN (transitional)
NLF
4,800
[0.1,0.85]
[1,10]
[−2,6]
50,000
Table 1: Composition of the UniFoil dataset [ 23 ] underlying the airfoil design case study of Section 7 .
Quantity
Value
Airfoil class / turbulence model
FT / SA
Mach / Re / AoA range used
[0.1, 0.85] / [1, 10] / [-2, 6]
Geometric design parameters ϑgeo. (dim. d1 )
14 coefficients
Performance labels y(ϑ)
(CL,CD,CM)
Transolver training / validation
325,485 / 17,130 (95% / 5%)
Diag-CFM training / validation
308,353 / 34,262 ( ≈ 90% / ≈ 10%)
Table 2: Design space and performance labels selected from UniFoil for the experiments in Section 7 .
Figure 2: RDC configuration: (a) CAD render illustrating the combustor geometry and pressure-sensor locations; (b) photograph of the experimental apparatus.
Figure 3: Experimental pressure data and reconstructed pressure field for a representative RDC operating condition.
Quantity
Value
Geometry
Outer radius / annulus width
45 / 10 mm
Centre-body length
60–130 mm
Outlet restriction
0, 25, 33, 50%
Operating conditions
Air mass flow rate
250–450 g/s
Table 3: Experimental design space and dataset composition for the RDC case study.
Quantity
Value
Data
Grid points N
512×512=262,144
Inputs / outputs per point
4+rref2=68 / 4
Training / held-out simulations
325,485 / 17,130
Architecture
Transolver blocks L
6
Table 4: Transolver configuration and training hyperparameters for the UniFoil case.
Quantity
Relative error
Velocity Ux
0.0478
Velocity Uy
0.187
Pressure coefficient Cp
0.257
Mach number Ma
0.273
Lift coefficient CL
0.399
Drag coefficient CD
1.21
Table 5: Transolver validation errors: mean per-sample relative L2 error over the fluid region and mean absolute relative error of force coefficients integrated from predicted and reference fields.
Figure 4: Velocity Ux field around a held-out airfoil at α=1.094∘ , Re=1.990×106 , Ma∞=0.598 : (a) interpolated ADflow solution, (b) Transolver prediction, (c) absolute error. The airfoil interior is masked. Reference and prediction share the same color scale.
Quantity
Value
Data
Conditioning parameters
4 ( L , OR, m˙air , ϕ )
Pressure-field resolution
16×256
Training / held-out RDC
170 / 42
Architecture
Latent distribution
N(0,I32)
Table 6: Conditional GAN architecture and training hyperparameters for phase-averaged RDC pressure-field generation.
Figure 5: Comparison of experimental and generated phase-averaged pressure fields for the held-out high-coherence case HC-04 ( L=100 mm, OR=33% , m˙air=353 g/s, ϕ=0.856 ). From left to right: experimental reference, mean of 1,000 generated fields, absolute difference, and standard deviation of the generated fields. Reference and prediction use the same pressure scale. Spline interpolation is applied for visualization only.
Figure 6: Extracted shock features for HC-04, shown for the experimental pressure map (left) and the generated ensemble mean (right). The primary and reflected peak trajectories and their linear fits yield the shock angles α and β . The horizontal line indicates the estimated height H and the cross marks the reflected-shock intercept yR . The same extraction method and pressure scale are used for both fields.
Figure 7: Comparison of extracted shock angles for held-out configurations at OR=25% and 33% . Blue markers denote the experimental fields and orange markers the cGAN ensemble means. The marker shapes indicate the equivalence ratio ( ϕ=0.70 : square, ϕ=0.85 : circle, ϕ=1.00 : triangle). The angles are extracted using the same procedure for experimental and generated pressure maps.
Model or selection
CL
CD
CM
Forward prediction: reference labels
Training mean
0.399
0.0334
0.0754
MLP ensemble
0.107
0.0132
0.0451
Diag-CFM
0.261±0.018
0.0193±0.0017
0.0788±0.0145
Round-trip error: surrogate predictions
All draws
0.125±0.004
0.0106±0.0001
0.0318±0.0013
Table 7: UniFoil RMSE. Diag-CFM and round-trip results show means ± standard deviations across five training seeds. We average round-trip errors over five sampling repetitions for each model. Zero-Deviation selects one of ten draws per target.
Model or selection rule
Speed
HDR
Forward prediction: measured labels
Training mean
147.6
0.1149
Forward reference
74.7
0.1110
Diag-CFM
79.6±1.2
0.1148±0.0015
Round-trip error: surrogate predictions
All draws
91.3±1.2
0.0958±0.0017
Table 8: RDC RMSE on 42 validation cases. Diag-CFM and round-trip results show means ± standard deviations across five training seeds. Zero-Deviation selects one of ten draws per target. Speed is in m/s, and HDR is dimensionless.
Generative artificial intelligence offers a new paradigm to design matter in high-dimensional spaces. However, its underlying mechanisms remain difficult to interpret and limit adoption in computational mechanics. This gap is striking because its core tools-diffusion, stochastic differential equations, and inverse problems-are fundamental to the mechanics of materials. Here we show that diffusion-based generative AI and computational mechanics are rooted in the same principles. We illustrate this connection using a three-ingredient burger as a minimal benchmark for material design in a low-dimensional space, where both forward and reverse diffusion admit analytical solutions: Markov chains with Bayesian inversion in the discrete case and the Ornstein-Uhlenbeck process with score-based reversal in the continuous case. We extend this framework to a high-dimensional design space with 146 ingredients and 8.9x10^43 possible configurations, where analytical solutions become intractable. We therefore learn the discrete and continuous reverse processes using neural network models that infer inverse dynamics from data. We train the models on only 2,260 recipes and generate one million samples that capture the statistical structure of the data, including ingredient prevalence and quantitative composition. We further generate five new burgers and validate them in a blinded restaurant-based sensory study with n = 101 participants, where three of the AI-designed burgers outperform the classical Big Mac in overall liking, flavor, and texture. These results establish diffusion-based generative modeling as a physically grounded approach to design in high-dimensional spaces. They position generative AI as a natural extension of computational mechanics, with applications from burgers to matter, and establish a path toward data-driven, physics-informed generative design.
Vahidullah Tac, Ellen Kuhl
Department of Mechanical Engineering, Stanford University, Stanford, California, United States
Large Language Models (LLMs) can generate Computer-Aided Design (CAD), yet lack physical comprehension required for reliable engineering design. Instead of attempting to implicitly learn physical laws from data, we propose a Hybrid Agentic-Physical Architecture that embeds validated knowledge-based engineering tools directly into the decision making loop of autonomous AI agents. In this framework, engineering design is formulated as a closed-loop, sequential decision making process guided by explicit physical verification. Based on a load case, dedicated agents iteratively plan, generate, evaluate, and revise engineering designs using knowledge-based tools as a feedback signal. We introduce a benchmark dataset and metrics for assessing functional validity in generative CAD. Our system generates more complex and physically verified designs, with a 4.2 increase in structural complexity and improving compile rate by 3.5% compared to similar agentic methods. The codebase, prompts and dataset will be made publicly available to support reproducibility and future research.
Elias Berger, Muhammad Usama, Jan Mehlstäubl +2
Dresden University of Technology, Dresden, Germany · MAN Truck & Bus SE, Munich, Germany · German Research Center for Artificial Intelligence, Kaiserslautern, Germany +1
The need to burn 100% H2 in high efficient gas turbines featuring low NOx combustion in premix mode require the complete redesign of the combustion system to ensure stable operation without any flashback. Since all engine frames featuring a power range from 4 MW up to 600 MW are affected, a huge design effort is expected. To reduce this effort, especially to transfer knowledge between the different engine classes, generative design methods using latest AI technology will provide promising potential. In this work, this challenge is approached utilizing the current advances in generative artificial intelligence. We train an Invertible Neural Network (INN) on an expandable database of geometrically parameterized combustor designs with simulated performance labels. Utilizing the INN in its inverse direction, multiple design proposals are generated which fulfill specified performance labels.
Patrick Krüger, Hanno Gottschalk, Werner Krebs +1
Department of Mathematics, Technische Universität Berlin, Straße des 17. Juni 135, Berlin, Berlin, Germany · Sienmens Energy, Rheinstraße 100, Mühlheim an der Ruhr, NRW, Germany