Deep learning algorithm produce best solutions to Schrödinger equations


A new deep learning algorithm by researchers at the University of Vienna, including the Department of Mathematics, provides more accurate numerical solutions to the Schrödinger equation than ever before for a range of different molecules with relatively little computational effort. Surprisingly, the researchers found that while some "pre-training" of the algorithm could improve its predictive abilities, more extensive training was actively detrimental. In order to find more accurate solutions for the wave functions of molecules more quickly, the researchers have developed deep learning algorithms that sample the energies at multiple points and teach themselves to search for the ground state of the molecule.

© Universität Wien/Gebhard Sengmüller