Masato SUMIDA’s team at RIKEN has used deep learning procedures (AI) to design organic molecules that absorb light at desired wavelengths. They first set the absorption wavelength to be 300 nm, 400 nm, 500 nm, 600 nm. Then, they chose structural formula of 13,000 organic molecules of about 400 m.w. constituted by hydrogen (H), carbon (C), nitrogen (N), and oxygen (O) and read out their electronic structures by the deep learning method “Recurrent Neural Network (RNN)”. Next, they searched molecules with five absorption wavelengths by “Monte Carlo Tree Search (MCTS)” and calculated their properties and stability by “Density Functional Theory (DFT)”, a molecular simulation technology based on quantum mechanics. Following these steps, they were able to find candidates of stable organic molecules with the desired absorption wavelength. The number of molecules generated by RNN and MCTS was 3,200, 86 of which were molecules with stable and desired absorption wavelength predicted by DFT. In addition, six of these 86 molecules were reported to have been synthesized in the past. After synthesizing these molecules and measuring UV absorption spectra, 5 out of 6 exhibited the desired absorption wavelength. An example is shown in the figure. The authors expect that their procedure speeds up the development of functional molecules in organic electronics fields such as light collecting materials for solar cells, electric storage materials, or light emitting materials.

RIKEN news release, August 24, 2018