Publications Repository - Helmholtz-Zentrum Dresden-Rossendorf
|Total number to be selected: 1 Title record|
Neural Solvers are neural network-based solvers for partial differential equations and inverse problems. The framework implements scalable physics-informed neural networks Physics-informed neural networks allow strong scaling by design. Therefore, we have developed a framework that uses data parallelism to accelerate the training of physics-informed neural networks significantly. To implement data parallelism, we use the Horovod framework, which provides near-ideal speedup on multi-GPU regimes.
Keywords: PINNs; PDEs; Neural Solver; Scalable AI
Software in the HZDR data repository RODARE
Publication date: 2021-09-06