Towards Building an AI-Integrated Computer- Aided Design Platform for Design Research Molla Hafizur Rahman, Zhenghui Sha (PI) Department of Mechanical Engineering, University of Arkansas, Fayetteville IMECE Track 16-1 NSF Research Poster Competition IMECE 2019-13335 SIDI LAB DISCOVER MORE Project Objectives and Goals • To develop a fine-grained data-driven research platform to support studies in engineering system design. • To identify beneficial design patterns and aid designer grouping based on their sequential design behaviors. • To computationally model and predict human sequential decisions using deep-learning methods. • The overall goal of this project is to build an AI-integrated computer-aided design (CAD) tool for data-driven design research. Future Work • To develop a bi-level framework using embedding technique to predict human design actions based on the current model which predicts design processes. • To integrate system thinking factors into the predictive models by extracting human psychological factors and cognitive skill. • To implement reinforcement learning, such as Markov Decision Process (MDP), to further advance the artificial intelligence in our platform. Conclusions and Contributions References [1] P. Berkhin, "A survey of clustering data mining techniques," in Grouping multidimensional data: Springer, 2006, pp. 25-71. [2] Gero, J. S., 1990. “Design prototypes: a knowledge representation schema for design”.AI magazine,11(4), p. 26. [3] Goodfellow, I., Y. Bengio, and A. Courville, Deep learning. 2016: MIT press Publications Experimental Setup • We developed an open-source research platform that facilitates fine-grained data-driven design thinking studies based on Energy3D. • We developed an approach based on Markov chain model and clustering algorithms to identify design patterns and designers of similar behaviors. • We developed a deep-learning based approach that achieves higher prediction accuracy than the traditional sequential models, e.g., hidden Markov model and Markov chain model in design field. • We improved our deep-learning models by integrating both human attributes and human design actions which shows promising prediction. • In summary, this project lays a stepping stone towards building an AI-integrated computer-aided design platform for design research. Background 53 54 55 56 57 58 59 60 61 62 63 64 57.29 58.26 59.27 56.73 60.6 59.24 60.91 59.59 60.88 63.38 61.98 60.99 ACCURACY (%) Energy-plus home dataset Parking lot dataset ….. Function-Behavior-Structure Based Design Process Model Add wall Add wall Edit window Remove window Add roof Energy annual analysis Formulation Formulation Synthesis Reformulation 1 ….. Formulation Analysis (b) Recurrent neural network [6] Acknowledgments We gratefully acknowledge the financial support from the National Science Foundation through grant # 1842588. [1] Rahman, M.H., et al. Automatic Clustering of Sequential Design Behaviors. in ASME 2018 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference. [2] Rahman, M.H., Schimpf, C., Xie, C. and Sha, Z., 2019. A Computer-Aided Design Based Research Platform for Design Thinking Studies. Journal of Mechanical Design, 141(12). [3] Rahman M.H. et al. “A Deep learning-based Approach to predict sequential design decision”, in ASME 2019 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference . (Robert E. Fulton Best Paper Award) Figure 2: Computational models used in this study • We transformed Energy3D, a solar system design software, to a research platform for data-driven design thinking studies [1]. • Energy3D has various unique features for system design research. Figure 1: Build an AI-Integrated Design Tool • Function-Behavior-Structure (FBS) based design process model is used to facilitate the abstraction and modeling of design thinking and help reduce the dimensionality in modeling. • Unsupervised cluster algorithms are used to identify designers of similar sequential behavior [4]. • Deep-learning models, particularly the recurrent neural network and its variant long short-term memory unit (LSTM) and gated recurrent unit (GRU), are adopted to predict future design decisions. Markov chain Clustering methods Final clusters One-hot vector Deep learning model Predicted actions K-means Hierarchical agglomerative Network-based LSTM RNN GRU Data and Results Improvised deep learning models for better prediction accuracy Direct model Indirect model We conducted two design experiments: Energy-Plus Home Design and Solarized Parking Lot Design 52 participants 41 participants 0 10 20 30 40 50 60 70 MM LSTM HMM FFN RANDOM REP MODEL 45.3 60.6 57.6 57.4 43.2 59.5 57.5 56.6 14.3 33.0 ACCURACY (%) Training accuracy Testing accuracy Figure 4 : a) Comparison between deep-learning modes and traditionally used sequential models [3] b) Comparison between improved deep-learning models and baseline models (a) (b) 2 design cases ~100 students ~500 design actions 4-fold cross validation Fine-grained Design Sequence Synthesis Synthesis Input design sequence Predicted design sequence (a) Function-Behavior-Structure model [5] Figure 3: Clusters of designers of similar sequential behavior [2] Energy3D