OmniArch: Building Foundation Model for Scientific Computing

Authors: Tianyu Chen, Haoyi Zhou, Ying Li, Hao Wang, Chonghan Gao, Rongye Shi, Shanghang Zhang, Jianxin Li

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present Omni Arch, the first prototype aiming at solving multi-scale and multi-physics scientific computing problems with physical alignment. We addressed all three challenges with one unified architecture. Its pre-training stage contains a Fourier Encoder-decoder fading out the disharmony across separated dimensions and a Transformer backbone integrating quantities through temporal dynamics, and the novel PDEAligner performs physics-informed fine-tuning under flexible conditions. As far as we know, we first conduct 1D-2D-3D united pre-training on the PDEBench, and it sets not only new performance benchmarks for 1D, 2D, and 3D PDEs but also demonstrates exceptional adaptability to new physics via in-context and zero-shot learning approaches, which supports realistic engineering applications and foresight physics discovery.
Researcher Affiliation Academia 1SKLCCSE, School of Computer Science and Engineering, Beihang University, Beijing, China 2SKLMIP, School of Computer Science, Peking University, Beijing, China 3School of Artificial Intelligence, Beihang University, Beijing, China.
Pseudocode No The paper describes the architecture and methodology in detail using prose and mathematical equations but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes We release our models base and large variants1, concurrently addressing 1D, 2D, and 3D PDEs. 1https://openi.pcl.ac.cn/cty315/Omni Arch
Open Datasets Yes We collect 1D, 2D, and 3D datasets from the public PDEBench and PDEArena. ... PDEBench (Takamoto et al., 2022), PDEArena (Gupta & Brandstetter, 2022)
Dataset Splits Yes We structured the PDEBench data into distinct training, validation, and testing subsets. For one-dimensional (1D) PDEs, the training dataset comprises a selection from the CFD-1D, Reac Diff, Advection, Burgers, and diff-sorp datasets. From these, we reserve a random 10% sample of trajectories as the in-domain test set for each respective PDE equation. ... In the two-dimensional (2D) PDE case, we allocate 90% of trajectories from the CFD, diff-react, NSincom, and shallow water datasets for training. The remaining 10% form the in-domain test set. ... For three-dimensional (3D) PDEs, 90% of trajectories from the CFD-3D dataset are utilized for training, with the remaining 10% serving as the in-domain test set.
Hardware Specification Yes Fine-tuning is performed on an A40 GPU cluster, which has 40Gi B of memory per device.
Software Dependencies No The paper mentions software components like "LLa MA model", "BERT model", and "albert-math model" but does not specify their version numbers.
Experiment Setup Yes In our training process, the following strategies or decisions were made: Pre/Post Norm: Pre-norm, Norm Type: RMS Norm Type, Architecture: Decoder-Only, Attention-Type: Multi-scaled Attention, Position Embedding: Ro PE, Casual Masking: True We only evaluate the loss on the T + 1 physical fileds prediction. Hidden Size: 1024, initializer_range: 0.02, intermediate_size: 4096, num_attention_heads: 16. ... Table 9: Detailed setting of hyperparameters in pre-training the base and large models. ... Table 10: Detailed Fine-tuning Settings: The table provides the learning rate, width, modes, and batch size for 1D, 2D, and 3D data.