文献类型：期刊文献

英文题名：ML-PINN: A memory-efficient physics-informed Mamba-LSTM network for fast and accurate PDE solving

作者：Gao, Yiming[1];Wang, Bing[1];Lu, Jingyi[1];Tian, Zhou[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China

年份：2025

卷号：655

外文期刊名：NEUROCOMPUTING

收录：;EI(收录号：20253619111854);WOS:【SCI-EXPANDED(收录号:WOS:001566980800003)】；

基金：This work was supported by the National Key R&D Program of China (2022YFB3304903) .

语种：英文

外文关键词：Physics-informed neural network; Mamba; LSTM; Sequence-to-sequence modeling; Memory-efficient deep learning

摘要：Physics-informed neural networks (PINNs) have emerged as a novel PDE solver and demonstrated significant potential. However, most existing models struggle to achieve satisfactory performance in both solution accuracy and computational efficiency, particularly in time-dependent modeling scenarios where extended input sequences are typically required to maintain precision. This inevitably leads to excessive GPU memory allocation and prolonged training durations. In this work, we develop a novel architecture integrating Mamba with Long Short-Term Memory networks (LSTM), hereby referred to as the Physics-informed Mamba-LSTM Neural Network (ML-PINN). By using shorter input sequences, ML-PINN is able to maintain high accuracy while achieving up to a 36 % reduction in GPU memory consumption and a 48 % decrease in training time compared to methods such as PINNMamba which employs state-space models and PINNsFormer (a Transformer-based PINN framework).