US 12,242,572 B1
Real time, compact, dynamic, transfer learning models
Srinivasa Mohan, Pune (IN); Aniruddha Mukhopadhyay, Pittsburgh, PA (US); Siddhartha Mukherjee, Pittsburgh, PA (US); and RaviKumar Devaki, Tokyo (JP)
Assigned to ANSYS, INC., Canonsburg, PA (US)
Filed by ANSYS, INC., Canonsburg, PA (US)
Filed on Dec. 1, 2021, as Appl. No. 17/457,168.
Int. Cl. G06F 18/25 (2023.01); G06N 20/00 (2019.01); H04L 67/12 (2022.01)
CPC G06F 18/25 (2023.01) [G06N 20/00 (2019.01); H04L 67/12 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A non-transitory machine-readable medium having executable instructions to cause one or more processing units to perform a method of generating a machine learning fusion model for a physical system, the method comprising:
receiving low-fidelity input data and low-fidelity output data that represent a low-fidelity measurement of a physical system;
training a first model to predict the low-fidelity output data using the low-fidelity input data;
receiving high-fidelity input data and high-fidelity output data that represent a high fidelity measurement of the physical system;
invoking the first model with the high-fidelity input data to generate predicted low-fidelity data;
training a second model to predict the high-fidelity output data using the high-fidelity input data augmented with the predicted low-fidelity data, wherein an input to the second model includes the high-fidelity input data and the predicted low-fidelity data; and
creating a fusion model for the physical system based on the first model and the second model, the first model and the second model to receive input to the fusion model, the second model to receive output from the first model, and output of the fusion model corresponding to output of the second model.