Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization
Authors: Yuze Hao, Linchao Zhu, Yi Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our 3DID framework generates high-fidelity 3D models with greater versatility and superior performance on target objectives. |
| Researcher Affiliation | Academia | 1 College of Computer Science and Technology, Zhejiang University 2 The State Key Lab of Brain-Machine Intelligence, Zhejiang University |
| Pseudocode | No | The paper describes the methodology using mathematical equations and detailed textual descriptions in Sections 3.2, 3.3, and 3.4, but it does not include a clearly labeled pseudocode or algorithm block with structured steps. |
| Open Source Code | Yes | Our code will be open-sourced and available on Git Hub upon publication. |
| Open Datasets | Yes | We conduct our experiments on the Driv Aer Net++ dataset [72, 73], the largest available collection for aerodynamic car design, comprising over 8,000 diverse geometries paired with high-fidelity CFD simulations. ... Driv Aer Net++ [72]: CC BY-NC 4.0 |
| Dataset Splits | No | For training, we use the entire dataset. ... To train our model, we use the dataset with 8085 car designs to extract the point cloud and physical field. ... For supervision, we additionally sample another 50,000 points, each annotated with both occupancy values and physical field data. |
| Hardware Specification | Yes | We use four NVIDIA RTX A6000 GPUs to train the model. ... We use four NVIDIA RTX A6000 GPUs to train the diffusion model. ... We use two NVIDIA RTX A6000 GPUs to train the Mesh Graph Net. |
| Software Dependencies | Yes | Following Driv Aer Net++ [72], we employ the Open FOAM V11 [79] to conduct steady-state incompressible simulation using the k ω SST turbulence model, based on Menter s formulation [80]. ... Open FOAM [79]: GNU General Public License |
| Experiment Setup | Yes | We sample Ng = Np = 50,000 points for the geometry and physical field branches, respectively. The encoder consists of one cross-attention layer and eight self-attention layers, each with 12 attention heads and an embedding dimension of dz = 64. We use r = 64 for learnable tokens, and each with a channel dimension of de = 768, to enhance representation expressiveness. The latent code dimension is set to dz = 32. ... We optimize the VAE using a combination of three loss terms: binary cross-entropy loss (λBCE = 10 3), mean squared error for field regression (λMSE = 10 5), and KL divergence (λKL = 10 6). Training is performed using the Adam W optimizer [71] with a learning rate of 1 10 4, batch size 8 per GPU, for 100,000 steps. ... The diffusion process includes 1,000 denoising steps. ... The diffusion model is trained using a learning rate of 5 10 5, batch size of 4 per GPU, for 300,000 steps with the Adam W optimizer. ... we apply a Free-Form Deformation (FFD) grid with 20 6 6 control points along the x, y, and z axes, respectively. ... This model is trained with a learning rate of 1 10 5, batch size of 8 per GPU, for 100,000 steps, using Adam W as the optimizer. |