Embedded Files
Research Focus
My research lies at the intersection of Human-AI Interaction and Engineering Design, with a particular focus on developing advanced AI-powered design tools for designers and engineers.
Below, I have listed selected research projects I have worked on.
Research Projects
LLM Powered Metacognitive Support Agent for Engineering Design
Sept. 2025 – Present
Carnegie Mellon University
[Ongoing Project]
Develop a Large Language Model (LLM) powered proactive design agent for real-time metacognitive support.
Utilize the Self-Regulated Learning (SRL) theoretical framework along with a fine-tuned LLM for timely and effective intervention and design support.
Design and conduct human-subject experiments to validate the effectiveness of the proposed design agent.
Generative Design Agent for Lithium-ion Battery Pack Design
Apr. 2025 – Present
Carnegie Mellon University
[Ongoing Project]
Developed a self-contained generative AI design agent with physics-based simulation and design validation for lithium-ion battery pack design.
Improved and integrated numerical design optimization tools for battery performance and cost optimization.
Adapted and fine-tuned Large Language Models (LLMs) for intuitive and engaging user interaction.
Designed an LLM-based Retrieval-Augmented Generation (RAG) system for user design ideation and creativity augmentation.
Visualization for Design Decision Making Using Generative AI
May. 2024 – Apr. 2025
Carnegie Mellon University
Studied the influence of design visualization on design decision making when using generative AI-powered design tools in engineering design practices.
Examined the change in designers’ preferences and user biases introduced with visualized design rendering versus pure data-driven design approaches, through human-subject experiments.
Results suggested that providing only design performance data can lead to the best ability to select the most optimal design, while only seeing design rendering provides marginal help in determining optimal designs.
Guides future development and implementation of generative AI-powered engineering design tools.
Team Equal Contribution and Design Representation
May. 2024 – Feb. 2025
Carnegie Mellon University
Investigated the impact of design representation modalities on member equal contribution in engineering design teams.
Proposed a novel and comprehensive method for team equal contribution assessment.
Conducted multi-perspective quantitative data analyses to examine differences in equality of member contribution between experimental conditions.
Results suggested that the use of more complex modalities will not result in more inequality in member contribution in engineering design teams, guiding the proper use of design representation to facilitate productive and enjoyable engineering design practices.
Local-dependency for Machine Learning with Data Decomposition
Sept. 2023 – Apr. 2024
Carnegie Mellon University
Constructed a novel data decomposition method to strictly limit the scope of information for neural operators to predict the local properties, promoting better convergence and generalizability in engineering simulations.
Proposed a machine learning method that works coherently with finite element simulations to assist engineering analysis by obtaining a highly accurate and highly resolved solution field.
The method supports targeted computation of accurate physics properties based on localized super-resolution at selected points of interest.
The proposed new method showed significantly improved performance with accelerated training convergence and reduced computational cost in multiple physical phenomena datasets compared to previous models on large-scale engineering simulations, making it better suited for in-situ applications.
Team Shared Understanding with Design Representation
Jan. 2023 – Apr. 2024
Carnegie Mellon University
Investigated the relationship between using common design representations and the development of shared understanding in engineering design teams.
Designed and conducted human-subject experiments and comprehensive quantitative data analyses to unveil performance differences between experimental conditions from different data sources.
Performed qualitative data analysis through thematic analysis and provided qualitative insights for supporting quantitative findings.
The findings emphasized the importance of using visual representations in engineering design practices to achieve better team collaboration and performance, especially when the team is remotely distributed.
Human-AI Partnership for Engineering Design
Aug. 2022– May. 2023
Carnegie Mellon University
Explored the incorporation of AI in design teams to support team problem-solving and genuine human-AI partnership capable of mimicking the dynamic adaptability of humans to improve team performance and avoid adverse effects.
Integrated AI agents with both reactive and proactive behaviors in hybrid design teams to solve complex and evolving engineering design problems.
Team performance and problem-solving behaviors are examined using the HyForm collaborative research platform, simulating complex interdisciplinary design problems with unexpected changes in problem constraints midway through to simulate the nature of dynamically evolving engineering problems.
The findings demonstrated AI agents are capable of participating as active partners within human teams during large, complex engineering design tasks, showing promise for future integration in engineering design practices.
Stylized Facts Informed Neural Network for Stock Volatility Prediction
May. 2022 – Dec. 2022
Carnegie Mellon University
Studied stylized facts in the stock market and created related functions for model performance evaluations.
Examined simple LSTM (Long Short-Term Memory)-based neural networks for log-return prediction, log-return distribution prediction, and volatility prediction, and compared the performance with the well-established GARCH (Generalized Autoregressive Conditional Heteroskedasticity)-like statistical models.
Integrated the GARCH model and economic stylized facts to form novel regularization loss functions for stock volatility prediction, mitigating the problem of overfitting in previous ML models.
Constructed a new hybrid model that combines the advantages of LSTM neural networks and the GARCH model to achieve better performance in short-time stock volatility prediction than previous approaches.
Page updated
Report abuse