US 12,334,194 B2
Deep imitation learning for molecular inverse problems
Eric Jonas, Chicago, IL (US)
Assigned to The University of Chicago, Chicago, IL (US)
Appl. No. 17/773,647
Filed by University of Chicago, Chicago, IL (US)
PCT Filed Nov. 3, 2020, PCT No. PCT/US2020/058685
§ 371(c)(1), (2) Date May 2, 2022,
PCT Pub. No. WO2021/091883, PCT Pub. Date May 14, 2021.
Claims priority of provisional application 62/930,325, filed on Nov. 4, 2019.
Prior Publication US 2022/0383989 A1, Dec. 1, 2022
Int. Cl. G16C 20/30 (2019.01); G16C 20/20 (2019.01); G16C 20/70 (2019.01)
CPC G16C 20/30 (2019.02) [G16C 20/20 (2019.02); G16C 20/70 (2019.02)] 20 Claims
OG exemplary drawing
 
1. A method of determining a molecular structure of a compound, the method comprising:
obtaining a known molecular formula of the compound based on at least one of an observed spectrum and stoichiometric calculations;
determining edges that meet per-vertex constraints of the molecular formula by using a trained deep neural network to sequentially place new edges until a candidate structure is complete, wherein the trained deep neural network is trained with training pairs generated from known molecules with observed properties and deleted edges;
generating a plurality of candidate structures based on the determined edges;
evaluating the plurality of candidate structures; and
determining one candidate structure of the plurality of candidate structures as the molecular structure of the compound based on the evaluation of the plurality of candidate structures.