Hybrid Physical-Law Simulation With Multilayer Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The design of complex vehicles like aerial vehicles is hindered by the computational complexity and accuracy issues of physical models, particularly in computational fluid dynamics, which are too costly in terms of computing time and require approximations that compromise simulation accuracy.
Innovation Solution
A method that partially or totally substitutes physical equations of a subsystem with an artificial intelligence model, specifically using a multilayer neural network, coupled with modules based on behavioral physical laws, allowing for adaptive parameter updates through a learning phase to enhance simulation accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physical models based on Navier-Stokes equations are used for simulation, then simulation accuracy is improved, but computing time and computational cost increase significantly
Solution Approach 1:
The patent divides the simulation system into two parts: a neural network module that handles complex aerodynamic calculations and a traditional CFD module that handles other physical processes. This segmentation allows the computationally intensive neural network to be trained offline and reused, while the traditional CFD module provides physical accuracy where needed, resolving the contradiction between accuracy and computing time.
Solution Approach 2:
The patent creates a neural network model that copies and learns from the complex physical phenomena described by Navier-Stokes equations. Instead of directly solving the computationally intensive physical equations during simulation, the neural network provides a learned approximation that captures the essential aerodynamic behavior with much lower computational cost while maintaining accuracy.
2Loss of time
If approximations like RANS are used to reduce computational cost, then computing time is reduced, but simulation accuracy deteriorates
Solution Approach 1:
The patent replaces the traditional mechanical/mathematical physics-based RANS approximation system with a neural network system. The neural network is trained on high-fidelity data and can provide accurate aerodynamic predictions without relying on simplifying physical assumptions, thus achieving both low computational cost and high accuracy simultaneously.
Solution Approach 2:
The patent changes the fundamental parameter of the simulation approach from physical equations (Navier-Stokes, RANS) to neural network parameters (weights and biases). This parameter transformation allows the system to capture complex aerodynamic phenomena without the computational burden of solving the underlying physical equations, resolving the accuracy-computing time tradeoff.
3Loss of time
If machine learning models are used to substitute physical models, then computing time is reduced, but model accuracy and reliability become uncertain
Solution Approach 1:
The patent introduces a neural network as an intermediary between the input data and the simulation output. The neural network is trained on high-fidelity data from physical simulations or experiments, serving as a mediator that translates input parameters into accurate aerodynamic predictions without requiring real-time solution of physical equations, thus achieving both speed and reliability.
Solution Approach 2:
The patent implements a feedback mechanism where the neural network predictions are validated against physical models or experimental data. The neural network is trained using feedback from high-fidelity simulations, allowing it to learn and adapt to ensure accurate predictions. This feedback loop maintains reliability while enabling fast computation during actual simulations.
Data Source
AI summary
A method for simulating the behavior of a system (20) composed by a subsystem (10) in a physical environment (21), by a set of modules, each simulating a portion of said system, wherein at least one portion (11) of said subsystem (10) is simulated by a module implementing a multilayer neural network, the other modules implementing simulations based on behavioral physical laws; said method comprising the assessment (S2) of the results of the simulation to automatically decide (S3) on a strategy to trigger updates (S4) of parameters associated to said multilayer neural network.