Amortized Fourier Neural Operators

Authors: Zipeng Xiao, Siqi Kou, Hao Zhongkai, Bokai Lin, Zhijie Deng

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we validate the effectiveness of our proposed method by conducting extensive experiments on challenging benchmarks governed by typical solid and fluid PDEs.
Researcher Affiliation Academia 1 Qing Yuan Research Institute, SEIEE, Shanghai Jiao Tong University 2 Dept. of Comp. Sci. & Tech., Tsinghua University
Pseudocode No The paper does not contain any explicit pseudocode or algorithm blocks.
Open Source Code Yes Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes]
Open Datasets Yes We evaluate the performance of AM-FNO on six well-established benchmarks. These benchmarks include Burger, Darcy, and NS-2D, which are presented in regular grids with varying dimensions [18]. We extend our experiments to assess the method s performance in different geometries, including Pipe, Airfoil, and Elasticity benchmarks [17] Additionally, we incorporate the compressible fluid dynamics (CFD) 1D and 2D benchmarks [30].
Dataset Splits No Table 1 provides 'Ntrain' and 'Ntest' counts, but there is no explicit mention of a separate validation set or its split details.
Hardware Specification Yes The batch size is selected from {4, 8, 16, 32}, and the experiments are conducted on a single 4090 GPU.
Software Dependencies No The paper mentions software components like 'Adam W optimizer' and 'Gaussian Error Linear Unit (GELU)' but does not provide specific version numbers for any libraries or frameworks (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes We train all models for 500 epochs using the Adam W optimizer [23] with a cosine annealing scheduler [22]. The initial learning rate is 10 3, and the weight decay is set to 10 4. Our models consist of 4 layers with a width of 32 and process all the frequency modes of training data. The Gaussian Error Linear Unit (GELU) is used as the activation function [13]. For AM-FNO (KAN), the number of spline grids is selected from {24, 32, 48}, while for AM-FNO (MLP), the number of basis functions is set to 32 or 48. AM-FNO (MLP) utilizes Chebyshev basis functions as the orthogonal basis functions, as elaborated in the appendix. The batch size is selected from {4, 8, 16, 32}, and the experiments are conducted on a single 4090 GPU.