Production System Control Using Active Learning Design Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multi-disciplinary optimization tools for complex product design consume high computational resources and often rely on inaccurate surrogate models to predict performance, leading to inefficient optimization processes.
Innovation Solution
A method and system utilizing a trained machine learning module to generate predictive performance and constraint compliance distributions, allowing for the skipping of expensive simulations by predicting likely constraint violations or inferior performance, thereby reducing computational effort and focusing on promising design variants.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If surrogate models are used to predict simulation outcomes, then computing demand is reduced, but prediction accuracy deteriorates
Solution Approach 1:
The patent introduces an active learning framework that acts as an intermediary between full simulations and surrogate models. The active learning module selectively queries full simulations for specific design variants based on uncertainty sampling, using these queries to iteratively improve the surrogate model's accuracy. This mediator approach allows the system to maintain high prediction accuracy where needed while still benefiting from the computational efficiency of surrogate models for the majority of evaluations.
Solution Approach 2:
The system performs preliminary training of surrogate models using an initial dataset of full simulation results. This preliminary action creates a baseline predictive model that can quickly evaluate design variants. The active learning process then builds upon this foundation by strategically adding new training samples, progressively improving accuracy without requiring all possible simulations to be performed in advance.
2Measurement precision
If full simulations are run for all design variants, then prediction accuracy is maintained, but computational resources are excessively consumed
Solution Approach 1:
Instead of performing full simulations for all possible design variants (excessive action), the active learning framework performs simulations for only a carefully selected subset of design variants (partial action). The uncertainty sampling mechanism identifies specific regions of the design space where full simulations provide the most value for improving the surrogate model, avoiding unnecessary computations for design variants where the surrogate model is already sufficiently accurate.
Solution Approach 2:
The active learning system is self-adaptive, automatically identifying which design variants require full simulations based on the current state of the surrogate model's uncertainty. The system serves itself by using its own predictive outputs to guide the selection of new training samples, eliminating the need for external expert intervention to determine where computational resources should be allocated.
3Productivity
If the number of simulations is reduced using surrogate models, then computational effort is decreased, but optimization reliability deteriorates
Solution Approach 1:
The active learning framework incorporates continuous feedback loops where the performance and uncertainty of the surrogate model are evaluated after each iteration. Based on this feedback, the system adaptively selects new design variants for full simulation that target regions of highest uncertainty or potential improvement. This feedback mechanism ensures that the optimization process remains reliable by systematically addressing areas where the surrogate model's predictions are least certain.
Solution Approach 2:
The system dynamically adapts the surrogate model throughout the optimization process. Rather than using a static surrogate model trained once, the active learning framework continuously updates and retrained the surrogate model with new data from full simulations. This dynamic adaptation allows the system to maintain high reliability even as the optimization progresses and the design space is better understood.
Data Source
AI summary
A machine learning module is provided trained to generate from a design data record specifying a design variant, a predictive performance distribution and a constraint compliance distribution of the design variant. A predictive performance distribution and a constraint compliance distribution are generated by the machine learning module. The predictive performance distribution is compared with performance values of previously evaluated design data records. A simulation of the corresponding design variant is either run or skipped. A design evaluation record is output which includes a performance value and constraint compliance data each derived from the simulation if the simulation is run or, otherwise, each derived from the predictive performance distribution and the constraint compliance distribution. Depending on the design evaluation records, a performance-optimizing and constraint-compliant design data record is selected from the variety of design data records. The selected design data record is then output for controlling the production system.