M³: Reframing Training Measures for Discretized Physical Simulations
Researchers have introduced M³ (Multi-scale Morton Measure), a novel framework designed to enhance neural surrogate models for physical simulations. Traditional models trained on discretized samples often suffer from uneven supervision and spatial inconsistencies due to measure-induced bias. M³ addresses this by partitioning space based on physical variation and allocating supervision across multiple scales, ensuring balanced training measures. Tested on three industrial-scale datasets with diverse discretizations, the framework consistently improved prediction accuracy in continuous physical domains. Notably, it achieved up to 4.7 times lower error in large-scale volumetric cases. The method demonstrated robustness under aggressive subsampling, where models trained with M³ on reduced data points (1.6 million) outperformed those trained on higher-resolution data (160 million points). Specifically, it reduced physics-weighted relative L2 error by 3–4 times and mean squared error by up to 13 times. These findings underscore the critical role of data distribution in operator learning, positioning M³ as a scalable and data-efficient solution for achieving physically consistent modeling in artificial intelligence applications.
Wire timeline
M³: Reframing Training Measures for Discretized Physical Simulations
Researchers have introduced M³ (Multi-scale Morton Measure), a novel framework designed to enhance neural surrogate models for physical simulations. Traditional models trained on discretized samples often suffer from uneven supervision and spatial inconsistencies due to measure-induced bias. M³ addresses this by partitioning space based on physical variation and allocating supervision across multiple scales, ensuring balanced training measures. Tested on three industrial-scale datasets with diverse discretizations, the framework consistently improved prediction accuracy in continuous physical domains. Notably, it achieved up to 4.7 times lower error in large-scale volumetric cases. The method demonstrated robustness under aggressive subsampling, where models trained with M³ on reduced data points (1.6 million) outperformed those trained on higher-resolution data (160 million points). Specifically, it reduced physics-weighted relative L2 error by 3–4 times and mean squared error by up to 13 times. These findings underscore the critical role of data distribution in operator learning, positioning M³ as a scalable and data-efficient solution for achieving physically consistent modeling in artificial intelligence applications.
cs.AI updates on arXiv.org