New Framework Boosts Operator Learning for Large-Scale PDEs

Training operator-learning models for complex, large-scale problems governed by partial differential equations (PDEs) presents significant challenges, including high dimensionality, memory constraints, and limited data. These issues are prevalent in critical scientific and engineering fields such as climate modeling and geological carbon storage. Researchers have introduced a scalable framework called Karhunen-Loeve Deep Neural Network (KL-DNN) to address these limitations. This method constructs latent spaces using low-rank singular value decomposition and a nested Karhunen-Loeve expansion, enabling full-resolution predictions without data subsampling. Applied to geological carbon storage, KL-DNN demonstrated superior performance, achieving lower errors for pressure and CO2 saturation while offering a two-order-of-magnitude speedup compared to DeepONet. Its rapid training and inference times make it a practical solution for real-time decision support and uncertainty quantification in large-scale PDE problems.