A Physics-Based Virtual Environment for Enhancing the Quality of Deep Generative Designs

Abstract

In this paper, we present a method that uses a physics-based virtual environment to evaluate the feasibility of neural network-based generated designs. Deep learning models rely on large training data sets that are used for training. These training data sets are typically validated by human designers that have a conceptual understanding of the problem being solved. However, the requirement of human training data severely constrains the size and availability of training data for computer generated models due to the manual process of either creating or labeling such data sets. Furthermore, there may be misclassification errors that result from human labeling. To mitigate these challenges, we present a physics-based simulation environment that helps users discover correlations between the form of a generated design and the physical constraints that relate to its function. We hypothesize that training data that includes machine validated designs from a physics-based virtual environment will increase the probability of generative models creating functionally-feasible design concepts. A case study involving a generative model that is trained on over 70,000 human 2D boat sketches is used to test the hypothesis. Knowledge gained from testing this hypothesis will provide human designers with insights into the importance of training data in the resulting design solutions generated by deep neural networks.

Deep Generative Models in Engineering Design: A Review

“…Several other papers use the 2D shape synthesis domain to address the challenges of performance evaluation in DGMs. Dering et al [98], for example, apply an iterative retraining approach to boat sketches using an adaptation of the LSTM-based Sketch-RNN [138] as the generator. To evaluate candidate designs, performance is scored using a simulated environment in a game engine, within which the "behavior" (motion) of the design is learned.…”

“…The doodles were collected by Google from user-drawn sketches in an interactive sketching game. QuickDraw data is used in several works discussed [164,98]. Toh et al [165] introduce a dataset of 934 innovative milk frother design sketches with associated text descriptions that can potentially be used to train DGMs, perhaps factoring in Natural Language Processing (NLP) 12 .…”

A sparsity preserving genetic algorithm for extracting diverse functional 3D designs from deep generative neural networks

“…However, this approach only generates 2D images and leaves a human designer to determine how to realize the full 3D design in a functionally feasible way. Dering et al (2018) propose a method that leverages a VAE model called sketch-RNN to generate 2D designs that are then evaluated in a 2D simulation software. Instead of directly optimizing over the latent space, the authors continually replace designs in the data set with high performing designs from the evaluation software.…”

Spatial Grammar-Based Recurrent Neural Network for Design Form and Behavior Optimization

“…Evaluation of any design requires a model to assess its performance. To support low fidelity trade space analysis, the authors have actively been using game engines [17] (e.g., Unity3D 2 ) to provide a variety of physics-based capabilities including rigid bodies with mass properties, colliders, joints, custom meshes, particle systems, and scripting for customization. Recent implementations of RL plugins have been added to video game engines [18], creating intelligent agents that are readily extensible for solving problem, behavior, and design problems simultaneously.…”