DADO -- Low-Cost Query Strategies for Deep Active Design Optimization

TL;DR

This work applies deep active learning to the field of design optimization to reduce the number of computationally expensive numerical simulations widely used in industry and engineering and presents two query strategies for self-optimization to reduce the computational cost in multi-objective design optimization problems.

Abstract

In this experience report, we apply deep active learning to the field of design optimization to reduce the number of computationally expensive numerical simulations. We are interested in optimizing the design of structural components, where the shape is described by a set of parameters. If we can predict the performance based on these parameters and consider only the promising candidates for simulation, there is an enormous potential for saving computing power. We present two selection strategies for self-optimization to reduce the computational cost in multi-objective design optimization problems. Our proposed methodology provides an intuitive approach that is easy to apply, offers significant improvements over random sampling, and circumvents the need for uncertainty estimation. We evaluate our strategies on a large dataset from the domain of fluid dynamics and introduce two new evaluation metrics to determine the model's performance. Findings from our evaluation highlights the effectiveness of our selection strategies in accelerating design optimization. We believe that the introduced method is easily transferable to other self-optimization problems.

Figures (8)

• Figure 1: Parameterized geometry with 6 boundary points (defined by two design parameters each) in green and 16 curve parameters in red (adapted from decke2023dataset). The objective of this use-case is to find a U-Bend design with minimal pressure loss and high cooling power.
• Figure 2: Example of a certain design candidate solved to compute the pressure using a computationally expensive numerical simulation. The scalar target value pressure loss is derived from this result representation.
• Figure 3: The DAL process: The Expert Model is initially trained on $X\_train\_0$ and executes expensive numerical simulations. This corresponds to an omniscient oracle that provides the real performance of the candidates. The boxes on the left are only a visual aid to represent the whole parameter space (candidates in the boxes are randomly drawn from the whole space). In each iteration $i$, a random set of candidates $X\_draw\_i$ is drawn from the design space. The Surrogate Model predicts the performance for each candidate. Based on the predictions $y\_draw\_i$, the Selector chooses the best candidates $X\_aq\_i$ which are then used to improve the Surrogate Model. The expensive numerical simulation is thus only executed on already promising candidates. The cycle repeats until an abortion criterion is met, for instance, until a good-enough candidate is found or the budget is exhausted.
• Figure 4: Two proposed regression query strategies for multi-objective deep active design optimization shown using a multivariate Gaussian distribution.
• Figure 5: Evaluation of the low-budget experiment S1 with $aq\_size$ of 25. The rnd_MSE represents the MSE between $y\_draw\_i$ and its true annotations, while best_MSE denotes the MSE between $y\_aq\_i$ and its true annotations.