Active learning of deep surrogates for PDEs: application to metasurface design