Production System Control Using Predictive Design Variant Screening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current design optimization methods for complex products consume excessive computational resources due to high simulation demands, often relying on inaccurate surrogate models for performance prediction.
Innovation Solution
A method utilizing a trained machine learning module to generate predictive performance and constraint compliance distributions, allowing for the skipping of simulations where design variants are likely to violate constraints or perform poorly, thereby reducing computational effort while maintaining efficiency in identifying promising designs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If simulations are run for all design variants to ensure accurate performance evaluation, then measurement precision is improved, but computing time and computational resources are excessively consumed
Solution Approach 1:
The machine learning model is trained in advance on a subset of design variants to learn the relationship between design parameters and performance outcomes. This preliminary training enables the model to quickly predict performance for new design variants without requiring full simulations, thus reducing computing time while maintaining evaluation accuracy.
Solution Approach 2:
A machine learning model is introduced as an intermediary between design variant generation and performance evaluation. The model approximates the complex simulation process, providing accurate performance predictions without requiring actual simulations for every design variant, thereby resolving the contradiction between precision and computing time.
2Loss of energy
If surrogate models are used to predict simulation outcomes, then computing resources are reduced, but prediction accuracy deteriorates
Solution Approach 1:
The machine learning model is trained in advance on a carefully selected subset of design variants that capture the essential design space characteristics. This preliminary training with high-quality data ensures the model achieves accurate predictions while requiring minimal computational resources during the actual optimization process.
Solution Approach 2:
The approach changes the parameters of the predictive model by using a machine learning model with appropriate architecture and training data selection, rather than traditional surrogate models. This parameter change enables the model to achieve both low computational cost and high prediction accuracy simultaneously.
3Manufacturing precision
If all design variants are evaluated through simulation to identify optimal designs, then manufacturing precision is improved, but productivity is reduced due to excessive computational effort
Solution Approach 1:
The machine learning model performs preliminary evaluation of design variants by predicting their performance based on training data. This preliminary action filters out poor designs early in the process, allowing only promising variants to undergo full simulation evaluation, thereby maintaining design optimization accuracy while dramatically increasing evaluation throughput.
Solution Approach 2:
Instead of performing full simulation evaluation on all design variants, the method applies partial evaluation using the machine learning model for all variants and reserves full simulation only for the most promising candidates. This partial action approach maintains sufficient accuracy for decision-making while significantly improving productivity.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A machine learning module (BNN) is provided trained to generate from a design data record (DR) specifying a design variant, a predictive performance distribution (PD) and a constraint compliance distribution (CCD) of the design variant. Furthermore, a variety of design data records (DR) is generated, for each of which the following steps are performed: - A predictive performance distribution (PD) and a constraint compliance distribution (CCD) are generated by the machine learning module (BNN). - The predictive performance distribution (PD) is compared with performance values (PPV) of previously evaluated design data records. - Depending on the constraint compliance distribution (CCD) and the comparison, a simulation (SIM) of the corresponding design variant is either run or skipped. - A design evaluation record (DER) is output which comprises 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 (PD) and the constraint compliance distribution (CCD). Depending on the design evaluation records (DER), a performance-optimizing and constraint-compliant design data record (ODR) is selected from the variety of design data records (DR). The selected design data record (ODR) is then output for controlling the production system (PP).