Towards General Neural Surrogate Solvers with Specialized Neural Accelerators
Authors: Chenkai Mao, Robert Lupoiu, Tianxiang Dai, Mingkun Chen, Jonathan Fan
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We tailor SNAP-DDM to 2D electromagnetics and fluidic flow problems and show how innovations in network architecture and loss function engineering can produce specialized surrogate subdomain solvers with near unity accuracy. We utilize these solvers with standard DDM algorithms to accurately solve freeform electromagnetics and fluids problems featuring a wide range of domain sizes. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering, Stanford, Palo Alto, USA. |
| Pseudocode | No | The paper includes flowcharts and descriptions of procedures, but no explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for this project could be found at: https://github.com/Chenkai Mao97/SNAP-DDM |
| Open Datasets | No | The paper states: 'We used an established finite-difference frequency domain (FDFD) solver (Hughes et al., 2019) to generate 1000 fullwave simulations...'. It then crops this data to 'produce an 100k material dataset, an 1M material dataset...'. This is custom-generated data, and no public access link, DOI, or formal citation for this specific dataset is provided. |
| Dataset Splits | No | The paper mentions 'The 100k and 1M data are split into 90% training data and 10% test data.' However, it does not explicitly specify a validation split or how it was handled. |
| Hardware Specification | Yes | SNAP-DDM is run with one NVIDIA RTX A6000 GPU |
| Software Dependencies | No | The paper mentions software like 'fvcore', 'ceviche FDFD solver', and 'Adam optimizer', but it does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | All models use a batchsize of 64 and are trained for 100 epochs for 100k training data or 50 epochs for 1M training data. The Adam optimizer with an individually fine-tuned learning rate is used with an exponential decay learning rate scheduler that decreases the learning rate by 30 by the end of training. A padding of 20 grids is applied to all FNOs. ... All the FNO models are trained with α = 0.3 and the U-Net and Swin Transformer are trained with α = 0.1. |