US 12,450,468 B2
Physics augmented neural networks configured for operating in environments that mix order and chaos
William Lawrence Ditto, Wake Forest, NC (US); John Florian Lindner, Wooster, OH (US); Sudeshna Sinha, Punjab (IN); Scott Thomas Miller, Brooklyn, NY (US); Anshul Choudhary, Raleigh, NC (US); and Elliott Gregory Holliday, Durham, NC (US)
Assigned to NORTH CAROLINA STATE UNIVERSITY, Raleigh, NC (US)
Filed by North Carolina State University, Raleigh, NC (US)
Filed on Oct. 1, 2021, as Appl. No. 17/492,172.
Claims priority of provisional application 63/086,549, filed on Oct. 1, 2020.
Prior Publication US 2022/0108151 A1, Apr. 7, 2022
Int. Cl. G06N 3/088 (2023.01); G06N 3/04 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/0418 (2013.01) [G06N 3/045 (2023.01); G06N 3/088 (2013.01)] 30 Claims
OG exemplary drawing
 
1. A method for utilizing physics augmented neural networks configured for operating in environments that mix order and chaos, the method comprising:
utilizing a neural network (NN) pre-processor to convert noncanonical coordinates associated with a dynamical system to canonical coordinates, wherein the noncanonical coordinates include position and velocity coordinates and the canonical coordinates have values that are nonlinear functions of values of the position and velocity coordinates and include position and conjugate momenta coordinates;
concatenating a Hamiltonian neural network (HNN) to the NN pre-processor to create a generalized HNN, wherein concatenating the HNN to the NN pre-processor includes connecting an output of the NN pre-processor to an input of the HNN and connecting an input of the NN pre-processor to receive the noncanonical coordinates such that when the NN pre-processor receives and converts the noncanonical coordinates, the NN pre-processor outputs the canonical coordinates to the HNN;
training the generalized HNN to learn nonlinear dynamics present in the dynamical system from generic training data, wherein training the generalized HNN includes training the NN pre-processor to output the canonical coordinates given the noncanonical coordinates as input and training the HNN to output values produced by a Hamiltonian function of the canonical coordinates;
utilizing the trained generalized HNN to forecast the nonlinear dynamics, wherein utilizing the trained generalized HNN to forecast the nonlinear dynamics includes inputting values of the noncanonical coordinates to the trained NN pre-processor, generating, by the NN pre-processor, values of the canonical coordinates, and generating, by the HNN, values produced by the Hamiltonian function of the values of the canonical coordinates output from the NN pre-processor; and
quantifying chaotic behavior from the forecasted nonlinear dynamics to discover and map one or more transitions between orderly states and chaotic states exhibited by the dynamical system, wherein the dynamical system comprises a physical system and quantifying the chaotic behavior includes using the values output by the HNN to describe motion of the physical system.