Machine Controller Using Residual-Guided Physics-ML Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine controllers face challenges in achieving high accuracy and reducing computational effort for real-time simulations, with reduced order models leading to higher simulation errors and machine learning models being unreliable due to their black-box nature, while physics-informed neural networks require re-training and are not competitive for real-time simulations.
Innovation Solution
A machine controller is configured with a machine learning module and simulation module, using geometry and physical data to generate first physical property values, evaluate compatibility, and train the module to minimize residuals, allowing the machine learning module to predict second physical properties without full simulation, thus reducing computational effort and enhancing accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If reduced order models are used for real-time simulations, then computational resources are reduced, but simulation accuracy deteriorates
Solution Approach 1:
The system segments the simulation task into two parts: a reduced order model that provides fast computation and a correction module that restores accuracy. The corrected reduced order model applies correction terms based on the discrepancy between reduced order model predictions and full order model predictions, thereby maintaining high accuracy while preserving computational efficiency.
Solution Approach 2:
The system changes parameters by introducing correction terms that adjust the reduced order model outputs. These correction terms are derived from the difference between reduced order model predictions and full order model predictions, allowing the system to adapt the model accuracy dynamically while maintaining low computational cost.
2Speed
If surrogate models based on machine learning are used, then real-time simulation capability is achieved, but reliability deteriorates due to black-box nature
Solution Approach 1:
The full order model acts as an intermediary that validates and corrects the surrogate model predictions. When the surrogate model predicts states outside its training data coverage, the system uses the full order model to provide corrections, ensuring reliability while maintaining real-time performance through the primarily lightweight surrogate model.
Solution Approach 2:
The system implements feedback by continuously comparing surrogate model predictions with full order model predictions and using this discrepancy to correct the surrogate model outputs. This feedback mechanism ensures that even when the surrogate model operates in uncertain regions, the final predictions remain reliable through systematic correction.
3Reliability
If physics informed neural networks are used, then physical constraints are incorporated, but computational complexity and re-training requirements increase
Solution Approach 1:
Instead of fully implementing physics informed neural networks that require expensive re-training, the system applies partial physics-based correction only when needed. The correction is applied selectively based on the surrogate model's confidence in its predictions, achieving physical constraint adherence without the full computational overhead of complete physics-informed architecture.
Data Source
AI summary
A machine controller, geometry data and measured physical data of a machine is provided. The geometry data and the physical data are input to a machine learning module and to a simulation module of the machine controller. By the input data, the simulation module generates first values of a first physical property of a component of the machine on a discretized grid. Furthermore, an evaluator is provided for evaluating a physical compatibility of the first values with second values of a second physical property of the component, and for generating a residual quantifying the compatibility. The evaluator evaluates the compatibility of the first values with output data of the machine learning module and generates a resulting residual. Moreover, the machine learning module is trained to minimize the resulting residual, thus configuring the machine controller for controlling the machine by the output data of the trained machine learning module.