Francesco Ruscio

Generative models have transformed the synthesis of complex data, with remarkable success in domains such as image generation, and have already demonstrated major scientific impact through advances such as structure prediction in biology. This PhD project will investigate the fundamental principles of generative modeling with the aim of developing methods that extend beyond standard data generation to broader scientific settings. In particular, the research will focus on models that can produce fast and accurate samples, provide tractable likelihood estimates, and be steered at inference time to guide the search toward relevant regions of interest. A central objective is to enable scientific exploration beyond simple interpolation, allowing the generation of novel, out-of-distribution candidates in combination with suitable verification or evaluation mechanisms.

A first research direction will address efficient sampling and likelihood estimation, which are essential ingredients for applications such as Boltzmann generators, where one seeks to sample efficiently from complex physical distributions, as well as for adaptive inference-time control. Building on these capabilities, the project will develop generative modeling frameworks for guided exploration of previously uncharted scientific design spaces. The resulting methods will be applied in biological discovery, including protein engineering in collaboration with AITHYRA, and in materials science, with the goal of accelerating the discovery of novel materials. In the latter context, the project will also investigate whether generative models can emulate matter at density-functional-theory-level fidelity while substantially reducing the cost of conventional first-principles simulations.