Aligning Optimization Trajectories with Diffusion Models for Constrained Design Generation
Authors: Giorgio Giannone, Akash Srivastava, Ole Winther, Faez Ahmed
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our framework to structural topology optimization, a fundamental problem in mechanical design, evaluating its performance on in- and out-of-distribution configurations. Our results demonstrate that TA outperforms state-of-the-art deep generative models on in-distribution configurations and halves the inference computational cost. |
| Researcher Affiliation | Collaboration | Giorgio Giannone Massachusetts Institute of Technology Technical University of Denmark ggiorgio@mit.edu Akash Srivastava MIT-IBM Watson AI Lab akashsri@mit.edu Ole Winther Technical University of Denmark University of Copenhagen olwi@dtu.dk Faez Ahmed Massachusetts Institute of Technology faez@mit.edu |
| Pseudocode | Yes | C Algorithms Algorithm 1 DOM with Trajectory Alignment Require: Optimized Topologies X0 Require: Constraints C = (BC, L, V F) Require: Intermediate Optimization Steps Xopt while Training do Sample batch (x0, c, xopt) Compute Dense Relaxation k = K(bc, l) Compute Conditioning c = (k, c) Sample t, ϵ, xopt s(t) Compute xt q(xt|x0) Forward Model ϵθ(xt, c) Compute Loss Lt 1(x, c) = ||ϵθ(xt, c) ϵ||2 2 Trajectory Search xθ(xt, ϵθ) = (xt 1 αt ϵθ(xt, c))/ αt Trajectory Matching LTA = || xθ(xt, ϵθ) xopt s(t)||2 2 Compute Loss LDOM(θ) = Lt 1(x, c) + LTA Backpropagate θ θLDOM(θ) end while Algorithm 2 DOM without Trajectory Alignment |
| Open Source Code | Yes | Code: https://github.com/georgosgeorgos/trajectory-alignment-diffusion |
| Open Datasets | Yes | We use a dataset of optimized topologies gathered using SIMP as proposed in [66, 62]. ... We also release a large, multi-fidelity dataset of sub-optimal and optimal topologies obtained by solving minimum compliance optimization problems. This dataset contains low-resolution (64x64), high-resolution (256x256), optimal (120k), and suboptimal (600K) topologies. To our knowledge, this is the first large-scale dataset of optimized designs that also provides intermediate suboptimal iterations. |
| Dataset Splits | No | The paper states training on '30k optimized topologies' and testing on '1800 Test Configurations' but does not specify a validation split or how validation was performed. |
| Hardware Specification | No | The paper mentions that 'SIMP runs on CPU and the DDPM-based models on GPU' (Table 8 caption) but does not specify the models or other detailed hardware specifications for these components. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries or programming languages used in the experiments. |
| Experiment Setup | Yes | We train all the models for 200k steps on 30k optimized topologies on a 64x64 domain. ... We set the hyperparameters, conditioning structure, and training routine as proposed in [62]. ... Table 11: Relevant Hyperparameters for baselines and DOM on 64x64 datasets. ... Batch size 64 Architecture Unet Iterations 200K Learning rate 2e-4 Loss Lepsilon Optimizer Adam |