Composite Bayesian Networks for Inverse Manufacturing Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Integrated Computational Materials Engineering (ICME) faces challenges in predicting the relationships between a material's structure, processes, and properties, particularly in forward and inverse prediction problems within large design spaces, which hinders the efficient development of new materials and manufacturing processes.
Innovation Solution
A system and method utilizing a variant of the conditional Linear Gaussian Bayesian network to predict the configuration of manufacturing processes for desired properties, capable of representing non-linear relationships and learning piecewise linear approximations, by sampling composite models through Monte Carlo simulations and leveraging a knowledge base with ontological descriptions and Bayesian network models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If expensive simulations and experiments are used to explore the design space, then prediction accuracy is improved, but time consumption and cost increase
Solution Approach 1:
The patent segments the design space exploration into two stages: first using fast approximate data-driven models to narrow down the design space, then using expensive simulations and experiments only on the narrowed down portion. This segmentation allows achieving high prediction accuracy while significantly reducing time consumption by avoiding exhaustive exploration of the entire design space with costly methods.
Solution Approach 2:
The patent performs preliminary action by using fast data-driven models to pre-process and narrow down the design space before applying expensive simulations and experiments. This preliminary filtering action identifies promising regions of the design space, ensuring that subsequent costly computations are focused only on areas likely to yield valuable results, thereby reducing overall time consumption while maintaining prediction accuracy.
2Measurement precision
If expensive simulations and experiments are used to explore the design space, then prediction accuracy is improved, but computational cost increases
Solution Approach 1:
The patent segments the computational workload into two parts: initial fast data-driven modeling to identify promising design regions, followed by targeted expensive simulations only in those regions. This segmentation reduces computational cost by avoiding unnecessary expensive computations in regions unlikely to contain optimal solutions, while still achieving high prediction accuracy where it matters.
Solution Approach 2:
The patent performs preliminary computational action using efficient data-driven models to filter and narrow down the design space before committing computational resources to expensive simulations. This preliminary filtering reduces the volume of data requiring costly computational analysis, thereby reducing overall computational cost while preserving prediction accuracy in the regions of interest.
3Loss of information
If the entire design space is explored, then completeness of analysis is improved, but productivity decreases
Solution Approach 1:
The patent segments the design space exploration process into an initial broad screening phase using fast data-driven models, followed by a focused detailed analysis phase using expensive simulations only in promising regions. This segmentation maintains productivity by avoiding exhaustive exploration of the entire design space, while preserving completeness of analysis by ensuring that the most promising regions receive thorough investigation.
Solution Approach 2:
The patent performs preliminary analysis using fast data-driven models to identify and narrow down promising regions of the design space before conducting detailed analysis. This preliminary action maintains productivity by reducing the scope of subsequent expensive simulations to only those regions most likely to contain optimal solutions, while ensuring completeness by systematically identifying promising regions through the preliminary screening.
Data Source
Figure 1
Figure 2
Figure 3(a)
AI summary
The invention provides a system and method for inverse inference in a chain of manufacturing processes using Bayesian networks is provided. The method generates a composite Bayesian network model for a chain of manufacturing processes from Bayesian network models of the unit processes in the chain. The models of unit processes might have been learned independently in other contexts and stored in a knowledge repository. Models relevant for the current problem context are obtained from the knowledge repository and checked for compatibility using ontological information about their inputs and outputs. The obtained compatible Bayesian network models of unit processes are composed to generate a composite Bayesian network model for the chain. The generated composite Bayesian network model is sampled to perform inverse inference.