Machine Learning Module for Design Optimization with Uncertainty Quantification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current design optimization methods for complex products, such as robots and turbines, require extensive computational resources due to high simulation demands, often relying on inaccurate surrogate models that consume significant computational effort.
Innovation Solution
A machine learning module, trained to generate predictive performance signals and uncertainties, determines whether simulations are necessary, allowing for the skipping of expensive simulations when predictions are likely accurate, thereby reducing computational effort without compromising optimization accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If surrogate models based on machine learning are used to predict simulation outcomes, then computational resources are reduced, but prediction accuracy deteriorates
Solution Approach 1:
The patent introduces an uncertainty quantification intermediary that mediates between the machine learning prediction and the simulation decision. The uncertainty measure acts as a bridge, determining when predictions are reliable enough to replace simulations and when simulations are still necessary, thus resolving the accuracy-resource tradeoff
Solution Approach 2:
The patent changes the parameter space by not only predicting performance outcomes but also predicting uncertainty measures. This additional parameter (uncertainty) enables dynamic decision-making about simulation necessity, allowing the system to adaptively balance accuracy and computational cost
2Measurement precision
If simulations are run for all design variants to ensure accurate performance assessment, then optimization accuracy is maintained, but computational effort increases significantly
Solution Approach 1:
The patent applies partial action by running simulations only for a subset of design variants - specifically those where the machine learning uncertainty indicates potential inaccuracy. For variants with low uncertainty, predictions alone suffice, making the simulation process partial rather than exhaustive
Solution Approach 2:
The patent implements skipping by allowing the system to bypass simulations entirely for design variants where the uncertainty measure indicates high prediction confidence. This rushing through of uncertain cases while skipping confident ones optimizes the overall process efficiency
Data Source
AI summary
A machine learning module is provided which is trained to generate from a design data record specifying a design variant of a product, a first performance signal quantifying a predictive performance of the design variant and a predictive uncertainty of the predictive performance. A variety of design data records each specifying a design variant of the product is generated. For a respective design data record, the following steps are performed: a first performance signal and a corresponding predictive uncertainty are generated, depending on the predictive uncertainty, a simulation yielding a second performance signal quantifying a simulated performance of the corresponding design variant is either run or skipped, and a performance value is derived from the second performance signal if the simulation is run or, otherwise, from the first performance signal. Depending on the derived performance values, a performance-optimizing design data record is determined and output to control the production plant.