.png){kind=icon}
.png){kind=icon}
Return
Accelerating level-set based topology optimization using gradient and stochastic gradient online learning and prediction methods
.png){kind=icon}
PRE
AI
DOI：10.1007/s00158-025-04049-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper introduces two novel methods: gradient online learning and prediction (GoLap) and stochastic gradient online learning and prediction (SGoLap) for accelerating the topology optimization using level-set method (LSM). The GoLap and SGoLap methods utilize neural networks and recurrent neural networks to learn and predict topological derivatives online, reducing the need for repetitive FEA and sensitivity computations in the iterative optimization process. Using the new methods, the topological derivatives calculated and predicted in the previous and current iterations are used as training samples. Specifically, the GoLap and SGoLap use the full vectors and the stochastically selected subset of the topological derivatives as training samples respectively. By integrating these methods into LSMs using a reaction–diffusion equation, the total computational time is significantly reduced by up to 98.5% when solving the considered 2D and 3D topology optimization problems.
Keywords:
Topology optimization
Machine learning
Level-set method
Neural network
Recurrent neural network
Journal
IF:
4
Papers:
4.8K
Citations:
1.7W