Topology Optimization Acceleration via Machine Learning Surrogate Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional topology optimization methods are computationally expensive due to the need for solving multi-physics problems multiple times, and existing acceleration techniques have limited applicability as they rely on past data that may not transfer well to new optimization cases.
Innovation Solution
A machine learning-based framework that uses a fully connected deep neural network to predict sensitivity values, reducing the number of two-scale optimizations by learning from current iterations and continuously updating the model with new data, thereby accelerating the topology optimization process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional topology optimization methods are used to solve multi-physics problems multiple times, then accurate sensitivity evaluation is achieved, but computational cost and time consumption increase significantly
Solution Approach 1:
The patent creates a surrogate model (copy) of the expensive multi-physics simulation process. This surrogate model approximates the relationship between design variables and sensitivity values, allowing rapid evaluation without repeatedly solving the full multi-physics problem. The surrogate model is trained on a subset of data and then used to predict sensitivity values for new design configurations, dramatically reducing computational time while maintaining acceptable accuracy.
Solution Approach 2:
The patent performs preliminary action by pre-computing and storing sensitivity values from multi-physics simulations in a database before the actual optimization begins. This pre-computed data serves as training data for the surrogate model, enabling the optimization process to proceed without repeatedly solving the expensive multi-physics equations. The preliminary computation creates a knowledge base that accelerates subsequent optimization iterations.
2Productivity
If parallel computing or GPU-based programming is employed to reduce generation time, then computational speed improves, but system complexity and resource requirements increase
Solution Approach 1:
The patent substitutes the mechanical/computational approach of parallel processing with a data-driven approach using machine learning. Instead of using multiple processors or GPUs to solve the same expensive physics equations simultaneously, the method uses a trained surrogate model to predict results directly. This replaces the computational mechanics of parallel solving with statistical mechanics of prediction, reducing hardware complexity while maintaining or improving speed.
3Productivity
If past data from previous topology optimization runs is used to speed up new cases, then computational efficiency improves, but applicability is limited when new cases differ significantly from past cases
Solution Approach 1:
The patent makes the surrogate model dynamic by enabling online updates during the optimization process. As new design configurations are evaluated, their results are fed back to update and retrain the surrogate model, allowing it to adapt to new problem instances. This dynamic updating mechanism maintains the model's relevance and accuracy even when new optimization cases differ from the training data, bridging the gap between computational efficiency and adaptability.
Solution Approach 2:
The patent creates a universal surrogate modeling framework that can handle multiple types of topology optimization problems. By using a flexible machine learning architecture and online updating capability, the system serves multiple functions: it can be trained on diverse problem types, adapt to new cases, and provide rapid predictions across different optimization scenarios. This multi-functionality enhances both productivity and adaptability simultaneously.
Data Source
AI summary
A system and method for accelerating topology optimization of a design includes a topology optimization module configured to determine state variables of the topology using a two-scale topology optimization using design variables for a coarse-scale mesh and a fine-scale mesh for a number of optimization steps. A machine learning module includes a fully connected deep neural network having a tunable number of hidden layers configured to execute an initial training of a machine learning-based model using the history data, determine a predicted sensitivity value related to the design variables using the trained machine learning model, execute an online update of the machine learning-based model using updated history data, and update the design variables based on the predicted sensitivity value. The model predictions reduce the number of two-scale optimizations for each optimization step to occur only for initial training and for online model updates.