FermiNet: Quantum physics and chemistry from first principles

The recent publication by researchers at MIT proposes a solution to one of the most difficult challenges in computational quantum chemistry: computing excited states, which play a crucial role in matter-light interactions and can be significantly more challenging than prior methods. The authors demonstrate that their new method, known as FermiNet, is more robust and general compared to state-of-the-art methods by achieving mean absolute errors (MAEs) on excited states of around 0.1 eV. To achieve this, the researchers developed a novel approach for representing and computing ground states while also incorporating information from additional streams for materials exploration. The paper introduces Graph Networks for Materials Exploration (GNome), which is an AI tool designed to explore quantum scale challenges such as producing clean electricity or developing high temperature superconductors. The researchers believe that their new approach will help advance the field of computational quantum chemistry by providing a more robust and general way to compute excited states.