Converting infrared light to ultraviolet is one operation. Filtering an image in Fourier space is another. In conventional optical systems, you perform them sequentially: first convert the wavelength, then process the image. Two stages, two sets of optics, two alignment problems.

A doubly resonant metasurface performs both simultaneously.

The device supports two high-quality resonances — a toroidal dipole bound state and a magnetic dipole resonance — at different wavelengths. One enhances the infrared signal, the other enhances the pump beam. Four-wave mixing between the enhanced fields generates UV output. But the spatial structure of the pump beam is not uniform — it is deliberately patterned to select specific spatial frequencies. The nonlinear conversion and the Fourier filtering happen in the same physical process, because the spatial frequency content of the pump determines which spatial components of the signal are upconverted.

This is not integration in the engineering sense — two functions packaged into one device. It is integration in the physical sense — two operations that share the same nonlinear interaction. The pump's spatial structure acts as a transfer function applied during wavelength conversion, not after it. Change the pump pattern, and you change the filter. The image processing is reconfigurable without changing the metasurface.

The deeper point is that nonlinear optics is inherently a multi-field process. When multiple fields interact, their spatial, spectral, and temporal structures couple. Treating wavelength conversion as "just conversion" ignores the structural information carried by the auxiliary fields. Every nonlinear process is already performing computation on the spatial modes of its inputs — the question is whether you harness it or ignore it.

The metasurface does not add computational capability to a converter. It reveals the computation that was already there.