Hyperspectral imaging has become essential for many applications, enabling the acquisition of three-dimensional hyperspectral scenes that include two spatial dimensions and one spectral dimension. The spectral sampling of these scenes allows for precise analysis and discrimination of the compounds present in the scene.
However, these scenes contain vast amounts of data, often redundant.
Coded hyperspectral imaging (CASSI -- Coded Aperture Snapshot Spectral Imaging) addresses this by encoding this three-dimensional information into a two-dimensional acquisition using a coded aperture optical system.
This acquisition can contain all relevant information in a compact data volume, which a suitable algorithm can then exploit.

Here, we present the simulation of CASSI systems through an optical ray-tracing simulator that is both realistic and differentiable: FROMAGE (Fast Ray-tracing-renderer for Optimization of Modern Architectures with Gradient Evaluation).
This simulator leverages recent advances in data science and is fully differentiable, enabling the joint optimization of all system parameters, of the coded aperture mask and also of the processing algorithm by using gradient-descent-based schemes.

Using this simulator, we evaluated the impact of optical distortions and alignment errors on the processing of CASSI-type imagers.
Specifically, we demonstrated that, with state-of-the-art and custom reconstruction algorithms, the reconstructed information is highly robust to distortions and misalignments.

We also used the simulator to develop a prototype, from designing the system's dispersive elements to creating a digital twin.
We developed a proxy model for optical components whose design details are unknown.
This proxy model, based on a neural network, is designed to verify physical properties such as the rotational invariance of lenses or the reversibility of light.

Additionally, we developed an adaptive processing algorithm capable of reconstructing hyperspectral scenes from an adaptive number of acquisition-mask pairs.
In a critique of supervised learning -- which requires vast amounts of labeled data -- we further developed a self-supervised reconstruction algorithm based on two systems: a prototype and its digital twin, modeled in the simulator.

This work paves the way for more realistic CASSI simulations, with a simulator capable of precisely modeling, optimizing, and co-designing coded hyperspectral imagers.
We also aim to deepen established methods for hyperspectral data processing by developing new adaptive algorithms that use several acquisitions.