Machine learning for Design Optimization

Our recent work advances data-driven and optimization-based methodologies for the design and analysis of high-performance engineering systems, with a focus on accelerating design workflows while improving functional accuracy and robustness.

Sensor network design under uncertainty: Multifunctional structural materials possess attractive attributes that can be designed to realize smart system functionalities such as integrated sensing systems for failure diagnostics and prognostics. With the integrated sensing capabilities, real-time monitoring of potentially damaging structural responses becomes possible. However, due to various uncertainties introduced by structural material properties, manufacturing processes, as well as operating conditions, ensuring the robustness of sensing performance is of vital importance for smart sensing system development. This research presents a data-driven robust design approach to develop piezoelectric materials based structural sensing systems for failure diagnostics and prognostics. In the proposed approach, a detectability measure is defined to evaluate the performance of any given sensing system, and the sensoring system design problem can be formulated to maximize the detectability for different failure modes by optimally allocating piezoelectric materials into a target structure. This formulation can be conveniently solved using reliability-based design framework to ensure design robustness while considering the uncertainties.

Piezoelectric sensor design for joint monitoring (left); Experimental verification (middle); Smart wireless transmitter for battery free monitoring (right)

Related recent papers:

Neural Network-Based Surrogate Model in Postprocessing of Topology Optimized Structures Neural Computing and Applications, 2025 Jude Thaddeus Persia, Myung Kyun Sung, Soobum Lee, Devin E. Burns This paper presents a neural network–based surrogate modeling framework for efficient postprocessing of topology-optimized structures after CAD reconstruction. The proposed method accurately predicts stress responses across multiple loading conditions, significantly reducing the need for repeated finite element analyses. Journal | PDF

Custom Multi-Component Force Transducer Design Using Topology Optimization Engineering Research Express, 2025 Myung Kyun Sung, Soobum Lee, Devin Edward Burns, Jude Thaddeus Persia This study presents a topology optimization–based design methodology for custom multi-component force transducers. The proposed approach enables compact sensor designs with tailored sensitivity and minimal cross-axis coupling, making it suitable for advanced experimental and robotic applications. Journal | PDF

Reinforcement Learning–Based Delay Line Design for Crosstalk Minimization IEEE Access, 2024 Jaeho Jung, Younggyun Yu, Soobum Lee This paper introduces a reinforcement learning–based framework for the automated design of delay lines with minimized crosstalk. The proposed method enables efficient exploration of high-dimensional design spaces and demonstrates improved electromagnetic performance compared to conventional optimization approaches. Journal | PDF