Research map / Probabilistic and scientific modeling / Molecules and proteins
Deep Generative Model for Mechanical System Configuration Design
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GearFormer, a Transformer trained on a synthetic DSL-generated, physics-simulated gear train dataset to instantly generate configuration designs, and hybrid EDA/MCTS+Transformer search methods.
Problem
Mechanical system configuration design is a categorical combinatorial optimization with multiple requirements and black-box physics simulation, making search methods slow; no public dataset exists for gear train synthesis.
Approach
Treat gear train design as a DSL token sequence and train a conditional Transformer on requirement-to-sequence pairs (with a weight objective via Gumbel-Softmax), then use it alone or to complete sequences inside EDA and MCTS.
Key results
- GearFormer achieves 94.73% feasibility on the test set versus 0.48% for random generation.
- On 30 benchmark problems, EDA+GF and MCTS+GF improve over EDA and MCTS alone on position, speed, and motion vector metrics while using 1/10 of the candidate solutions (10e3 vs 10e4).
- GearFormer generates a candidate in 0.328 s, while a single search candidate evaluation takes 0.039 s, with a single EDA or MCTS run taking about 6.5 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace random grammar-conforming sequence generation with GearFormer | GearFormer test set · GearFormer test set | feasibility (%): 0.4766 -> 94.73 | Helped |
| Replace random grammar-conforming sequence generation with GearFormer | GearFormer test set · GearFormer test set | output position match (m): 0.860 -> 0.041 | Helped |
| Replace random grammar-conforming sequence generation with GearFormer | GearFormer test set · GearFormer test set | speed ratio match (log): 1.9393 -> 0.0229 | Helped |
| Replace random grammar-conforming sequence generation with GearFormer | GearFormer test set · GearFormer test set | output motion vector match (%): 16.96 -> 99.04 | Helped |
| Add GearFormer completion to EDA search | 30 random benchmark problems · 30 random benchmark problems | speed ratio match (log): 0.0930^{\pm0.132} -> 0.0098^{\pm0.016} | Helped |
| Add GearFormer completion to MCTS search | 30 random benchmark problems · 30 random benchmark problems | output position match (m): 0.760^{\pm0.400} -> 0.209^{\pm0.362} | Helped |
| Add GearFormer completion to EDA search | 30 random benchmark problems · 30 random benchmark problems | weight (kg): 1.98^{\pm1.55} -> 3.26^{\pm3.59} | Hurt |
| Add GearFormer completion to MCTS search | 30 random benchmark problems · 30 random benchmark problems | weight (kg): 3.20^{\pm3.09} -> 2.62^{\pm2.19} | Helped |
| Increase the objective-loss weight w1 (ablation) | GearFormer · Validation set | average weight | Helped |
| Increase the objective-loss weight w1 (ablation) | GearFormer · Validation set | validity and feasibility | Hurt |
Surprising findings
- Pure EDA and MCTS achieve lower average weight (1.98 kg and 3.20 kg) than GearFormer (5.98 kg), reflecting that search optimizes weight at the expense of requirement satisfaction.
Limitations
- Sequences limited to a maximum of 10 components and a restricted DSL and parts catalogue.
- Validation and test evaluation requires slow simulator calls, so only small validation/test sets were used.
- Search baselines were evaluated on only 30 sampled problems due to cost.
Experimental setup
Datasets: GearFormer dataset (synthetic, generated via DSL, parts catalogue, and physics simulator)
Benchmarks: GearFormer test set (synthetic) 30 randomly sampled benchmark problems from test set
Metrics: validity feasibility output position error speed ratio RMSLE output motion vector match input/output motion type match weight (kg) candidate count
Baselines: Random sequence generation EDA MCTS
Builds on: Transformer Gumbel-Softmax Estimation of distribution algorithm Monte Carlo tree search
Closest papers by meaning
- BikeBench: A Bicycle Design Benchmark for Generative Models with Objectives and Constraints (NeurIPS 2025 · similarity 0.86)
- GenCO: Generating Diverse Designs with Combinatorial Constraints (ICML 2024 · similarity 0.83)
- MorphVAE: Advancing Morphological Design of Voxel-Based Soft Robots with Variational Autoencoders (AAAI 2024 · similarity 0.83)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.83)
- Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering with Generative Optimization (NeurIPS 2026 · similarity 0.83)
- Compositional Generative Inverse Design (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.