Physical Field Control Using FEM Lagrangian Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques for system control in energy systems and infrastructure systems face challenges in real-time control due to the time-consuming nature of physical phenomenon simulations, such as fluid simulations and non-linear mathematical models, which require several hours to estimate physical fields.
Innovation Solution
The implementation of a control apparatus using finite element model-based Lagrangian networks (FEMLN) for rapid and accurate conversion of sensing data into temporally and spatially distributed physical fields, enabling real-time system control by discretizing objects and employing machine learning methods like Lagrangian neural networks for energy functional calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physical phenomenon simulation (fluid simulation, fluid-structure interaction simulation) is used to estimate physical field, then measurement precision is improved, but loss of time increases significantly (several hours required)
Solution Approach 1:
The patent creates a digital twin (virtual model) that copies the physical system's behavior. Instead of running time-consuming physical simulations, the digital twin is trained once using simulation data and then used for rapid real-time predictions. This copying approach maintains measurement precision while dramatically reducing loss of time during actual control operations.
Solution Approach 2:
The patent performs preliminary training of the digital twin model using simulation data before actual control operations. This preliminary action prepares the model in advance so that during real-time control, only fast inference is needed rather than full simulations. The heavy computational work is done beforehand, enabling real-time responses.
2Reliability
If conventional simulation methods are used for system control, then reliability of physical field estimation is improved, but productivity decreases due to inability to perform real-time control
Solution Approach 1:
The patent replaces the mechanical simulation computation system with a machine learning-based digital twin system. The traditional approach uses physics-based simulations (mechanical computation), while the invention substitutes this with a trained neural network model that performs rapid inference. This substitution maintains reliability through proper training while enabling real-time control for improved productivity.
Solution Approach 2:
The patent changes the computational parameters from solving differential equations in real-time to using a pre-trained model with fixed weights for rapid prediction. By transforming the problem from continuous simulation to discrete prediction based on learned patterns, the system achieves both reliability (through accurate training) and real-time performance (through fast inference).
Data Source
AI summary
A control apparatus according to an embodiment includes one or more hardware processors. The processors acquire, for each of elements, pieces of input data representing first physical field of a corresponding one of the elements at a first time point. The elements are obtained by discretization of an object to be controlled. The processors calculate, for each element, second physical field of a corresponding element at a second time point after the first time point. The second physical field is calculated based on a value of an energy functional representing energy of the corresponding element. The value of the energy functional is obtained by inputting the pieces of input data into an estimation model. The processors control the object such that control quantity based on the second physical field becomes a target value.