Topology optimization with physics-informed neural networks: application to noninvasive detection of hidden geometries