ML Differentiable Representation for Material and Subassembly Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing design optimization methods for physical structures struggle to integrate the optimization of geometric properties, material selection, and subassembly choices in a way that respects the limitations of real materials and product offerings, particularly due to the discrete nature of materials and subassemblies.
Innovation Solution
The use of machine learning techniques to create a differentiable representation of materials and subassemblies based on a training set of material and subassembly catalogs, allowing for simultaneous optimization of geometric, material, and subassembly parameters within a continuous surface described by geometric parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If classic gradient-based optimization is used for geometric parameters, then optimization efficiency is improved, but it cannot be applied to discrete material and subassembly parameters
Solution Approach 1:
The patent transforms discrete material and subassembly parameters into continuous parameters through learned continuous representations. Machine learning models are trained to map discrete catalog items to continuous parameter spaces, enabling gradient-based optimization methods to be applied uniformly across geometric, material, and subassembly parameters without losing the discrete nature of the final selections.
2Productivity
If material parameters are treated as continuous variables, then optimization can be performed using gradient-based methods, but the results may not match existing materials
Solution Approach 1:
The patent introduces machine learning models as intermediaries between the continuous optimization process and discrete material catalogs. These models learn the mapping from continuous parameter spaces to discrete material selections, ensuring that optimization results correspond to actual available materials and subassemblies rather than theoretical continuous values.
Solution Approach 2:
The patent creates learned continuous representations that copy or approximate the discrete material and subassembly catalogs in continuous space. These representations preserve the essential characteristics of real materials while enabling continuous optimization, allowing the system to work with simplified continuous models that accurately reflect discrete reality.
3Device complexity
If geometric parameters, material parameters, and subassembly parameters are optimized independently, then each optimization can be performed separately, but the interdependencies between these parameters are not captured
Solution Approach 1:
The patent merges the optimization of geometric, material, and subassembly parameters into a unified framework. By representing all three types of parameters in continuous spaces and using a single objective function, the system simultaneously optimizes all parameters together, capturing their interdependencies and achieving truly integrated design optimization rather than separate sequential optimizations.
Data Source
AI summary
A system for optimizing physical designs provides integrated optimization of design geometry, design materials, and design subassemblies by mapping a catalog of actual or available construction materials and subassemblies to a differentiable representation tractable for computerized optimization. New subassemblies may be generated by using the differential representation in conjunction with a decoder trained on the actual or available subassemblies.