The effective design of instruments that rely on the interaction of radiation with matter for their operation is a complex task. A full optimization of the many parameters involved may still be sought by leveraging recent progress in computer science. Key to such a goal is the definition of a utility function that models the true goals of the instrument. Such a function must account for the interplay between physical processes that are intrinsically stochastic in nature and the vast space of possible choices for the physical characteristics of the instrument. The construction of a differentiable model of all the ingredients of the information-extraction procedures, including data collection, detector response, pattern recognition, and all existing constraints, then allows the automatic exploration of the vast space of design choices and the search for their best combination. In this document we succinctly describe the research program of the MODE Collaboration (an acronym for Machine-learning Optimized Design of Experiments), which aims at developing tools based on deep learning techniques to achieve end-to-end optimization of the design of instruments via a fully differentiable pipeline capable of exploring the Pareto-optimal frontier of the utility function. The goal of MODE is to demonstrate those techniques on small-scale applications such as muon tomography or hadron therapy, to then gradually adapt them to the more ambitious task of exploring innovative solutions to the design of detectors for future particle collider experiments.