Generative FlowNets for Material Physics Discovery

In many real world problems, the design process involves iterative steps with generation of candidates and evaluation of these generated candidates to evaluate their utility for the relevant task. This evaluation can be done with say expensive wet lab experiments or some computational oracles. We might want to make the best use of both of these evaluations. So ideally, during generation we would like to jointly generate the candidate as well as the fidelity of the oracle to evaluate the candidate with.
GFlowNets is a neural net that models distributions over data structures like graphs, to sample from them as well as to estimate all kinds of probabilistic quantities (like free energies, conditional probabilities on arbitrary subsets of variables, or partition functions) which otherwise look intractable.
Nikita’s project aims to extend the applications of GFlowNets in the context of material discovery, tparticularly new solid-state superionic conductors. The project focuses on exploring multi-fidelity active learning, in which at each step of an optimization algorithm, the model selects both a candidate and the precision level of the oracle from which obtaining the annotation of the candidate.

Faculty Supervisor:

Yoshua Bengio

Student:

Partner:

Birla Institute of Technology and Science, Pilani

Discipline:

Computer science

Sector:

Artificial Intelligence; Life Sciences (not health); Green/Alternative Energy

University:

Université de Montréal

Program: