Selected Works:

“Machine Learning-Enabled Prediction of Spatiotemporal Strain Evolution in Architected Hydrogels”, ACS Applied Engineering Materials, 2026, 4, 5, 2718–2725. (DOI)

“Time-Resolved Regression Digital Twin for Continuous Inference of Encoded Mechanical Heterogeneity in Architected Materials,” ACS Applied Engineering Materials, 2026. (DOI)

T. Liu, G. Manogharan, and H. Sun*, “Label Scarcity Reveals Architecture-Dependent Robustness in CNN-Based Casting Defect Detection,” 2026, under review. (preprint)

“Closed-Loop Additive Manufacturing of Architected Energy Materials: Integrating Machine Learning and Recycling for Intelligent and Sustainable Design”, 2026, under review.

“Learning Transferable Mechanical Representations from Full-Field Deformation Patterns in Architected Hydrogels,” 2026, under review.

“Machine Learning Reveals Hidden Mechanical States Governing Adaptive Deformation Accommodation in Thick Composite Electrodes,” 2026, under review.

“From deformation evolution to material design: A spatiotemporal digital twin for architected systems,” manuscript in preparation, 2026.

Current Projects

Data-Efficient Machine Learning for Identifying Manufacturing- and Cycling-Induced Defects and Degradation in Rechargeable Batteries

Personnel: Ta-Wei Wang (Ph.D. student)

We developed a data-efficient machine learning framework to identify manufacturing-induced degradation regimes in thick structural battery electrodes. Rather than relying on long-term cycling to assess stability, the approach extracts early-cycle electrochemical signatures that encode the onset of chemo-mechanical instability. Controlled manufacturing variations, including distinct processing pathways under identical material composition and geometry, are used to induce separable degradation trajectories within 80 cycles. From the first 10 cycles, features such as capacity decay slope, coulombic efficiency variance, voltage hysteresis growth, and polarization evolution are systematically extracted to characterize early-stage behavior. Supervised learning models link these early indicators to later-stage retention and degradation classes, enabling prediction of instability onset before substantial performance loss occurs. Feature-importance analysis further reveals which electrochemical signatures are most sensitive to manufacturing-induced microstructural differences, providing physical interpretability beyond black-box prediction.

T. Wang, G. Manogharan, and H. Sun*, “Closed-Loop Additive Manufacturing of Architected Energy Materials: Integrating Machine Learning and Recycling for Intelligent and Sustainable Design”, 2026, under review.

H. Sun et al., “Multifunctional Gel Boundary Phases Enable Adaptive Deformation in Thick Battery Electrodes Revealed by a Mechanics-Informed Digital Twin,” 2026, under review.

Deep Learning for Casting Defect Detection

Personnel: Ta-Wei Wang (Ph.D. student)

We developed a deep learning framework for automated casting defect detection under progressively limited labeled data. Rather than evaluating model performance only under fully supervised conditions, the approach systematically reduces the available labeled training data to reveal how different convolutional neural network architectures respond to label scarcity. Although different architectures achieve comparable performance with abundant labeled data, their behavior diverges substantially as supervision becomes limited, revealing architecture-dependent robustness that cannot be inferred from full-data accuracy alone. Class-specific error analysis and Grad-CAM-based attention analysis further reveal how severe label scarcity alters prediction behavior, spatial attention, and model confidence. This framework provides practical guidance for selecting robust and interpretable AI models for casting defect detection and data-constrained manufacturing quality inspection.

T. Liu, G. Manogharan, and H. Sun*, “Label Scarcity Reveals Architecture-Dependent Robustness in CNN-Based Casting Defect Detection,” 2026, under review. (preprint)

AI-Driven 4D Printing of Shape Morphing Actuators

Personnel: Juchen Zhang (Ph.D. student)

We developed an AI-driven inverse design framework that integrates finite element (FE) simulation, generative machine learning, and halftone-based 4D printing. Rather than relying on trial-and-error fabrication, the system automatically computes the optimal halftone parameters needed to deliver a target actuation behavior. We construct large training datasets using FE simulations of halftone-encoded heterostructures with varied pattern distributions. Time-resolved strain mapping captures mechanical responses at both local and global scales. Surrogate forward neural networks and generative inverse models link user-defined mechanical targets to manufacturable design parameters, enabling rapid prediction and fabrication of actuators with tailored properties. In addition to mechanical tuning, halftone-regulated growth deformations, such as thermoresponsive shrinkage, are used to program 3D shape transformations. A deep residual neural network, trained on FE-generated deformation datasets, performs both forward shape prediction and inverse mapping to generate halftone patterns for a prescribed final geometry, without requiring complex derivative calculations.

H. Sun*, S. Traczik, V. Devine-Ducharme, “Machine Learning-Enabled Prediction of Spatiotemporal Strain Evolution in Architected Hydrogels,” ACS Applied Engineering Materials, 2026, 4, 5, 2718–2725. (DOI)

T. Liu, H. Yang, H. Sun*, “Learning Transferable Mechanical Representations from Full-Field Deformation Patterns in Architected Hydrogels, ” 2026, under review.

Digital-Twin-Driven Codesign and 4D Printing of Intelligent Systems

Personnel: Moonsung Park (Ph.D. student)

Nature provides examples of soft structures that can stretch, deform, heal, and transform shape. Inspired by these systems, this project develops a new physical–digital-twin framework to design and 4D print multifunctional soft materials, including hydrogels, shape-memory polymers, and elastomers with programmable mechanical strength and shape-morphing behaviors. Using dynamic projection grayscale lithography, we 4D print thin films that contain engineered cellular domains microscale patterns that locally tune stiffness, strain distribution, and deformation. These printed “physical twins” are paired with digital twins, multiphysics simulation, machine learning models, and data driven inverse design, to predict how the printed material will stretch, bend, morph into 3D shapes, or respond to heat, light, or solvents. By letting physical twins and digital twins continuously inform each other, the project moves beyond trial-and-error manufacturing toward predictive, uncertainty aware design. We aim to create soft material systems that can adapt shape, enhance mechanical performance, program sequential motion for soft robotics and actuation, and enable generalizable design rules for future autonomous 4D-printing platforms. This research integrates advanced manufacturing, machine learning, and multiphysics simulations into. digital-physical-twin-guided frameworks for applications in soft robotic grippers, adaptive biomedical devices, stretchable electronics, reconfigurable infrastructure, and any intelligent systems that change with external stimuli.

M. Park, T. Liu, and H. Sun*, “Machine Learning-Enabled Probabilistic Digital Twin for Halftone-Encoded Programmable Mechanical Heterogeneity,” 2026, under review.

M. Park, T. Liu, and H. Sun*, “Time-Resolved Regression Digital Twin for Continuous Inference of Encoded Mechanical Heterogeneity in Architected Materials,” ACS Applied Engineering Materials, 2026. (DOI)