Dynamic Reduced Normalized Gaussian Network Method for Topology Optimization of Electromagnetic Devices

This paper proposes an efficient method to represent topology using fewer variables by combining the normalized Gaussian network with a truncated eigendecomposition. In this method, new orthogonal basis functions are derived from the original Gaussian functions, allowing the main features of the topology to be captured compactly. Since the bases are orthogonal after decomposition, we also introduce a strategy to dynamically increase the number of basis functions during optimization, ensuring that each newly added basis does not interfere with the existing ones. By progressively adding bases, the topology evolves from a coarse representation to a more refined structure, which helps reduce optimization time and avoid getting trapped in poor local solutions.