Research

Jichuan's research sits at the intersection of structural dynamics, scientific machine learning, and structural health monitoring. Recurring threads include neural-operator surrogates for full-field response prediction, conditional diffusion for identifying unmeasured excitations from sparse sensors, multi-fidelity learning for systems where high-fidelity simulation is expensive, and reinforcement-learning controllers for vibration mitigation and real-time hybrid simulation.

InvSHMDiff: ground motion identification from sparse structural responses

• Conditional diffusion
• Inverse dynamics
• Virtual sensing

Ground motion is often unmeasured, while structural monitoring networks provide only sparse and sometimes noisy acceleration records. Recovering the excitation from those responses is an ill-posed inverse problem, especially when entire sensor channels are unavailable.

InvSHMDiff combines a Vision Transformer sensor encoder with a conditional diffusion model to generate plausible ground-acceleration histories from incomplete bridge responses. On a simulated cable-stayed bridge, mean R² exceeds 99.79% with half of the 26 sensors missing. The model also reconstructs 76 recorded earthquake inputs from simulated bridge responses; its excitation ensembles provide relative uncertainty indicators and support virtual sensing through a calibrated forward model. The study is published in Advanced Engineering Informatics, vol. 77, article 105330 (2027). See the published paper and the full method diagram.

Surrogate models for full-field cable-stayed bridge dynamics

• Neural operators
• DeepONets
• Structural dynamics

High-fidelity finite-element simulations resolve full-field bridge dynamics, but their cost limits design iteration, uncertainty quantification, and real-time decision support. Conventional surrogates often reduce the response to a few scalar quantities, losing spatial information that is important for engineering interpretation.

We developed a spatio-temporal full-field Extended DeepONet that predicts multiple dynamical fields in a single forward pass while preserving correlations among spatial outputs. Across three DeepONet variants evaluated on a cable-stayed bridge model, the full-field model provides the best accuracy and lowest inference cost. The study is accepted in the ASCE Journal of Computing in Civil Engineering (2026) and is available as an arXiv preprint.

Multi-fidelity surrogates for shipboard shock response

• Multi-fidelity learning
• Transfer learning
• Shock dynamics

High-fidelity shock simulations of shipboard structures capture whipping and resonant response, but their cost rules out the parameter sweeps that designers need. Low-fidelity models are cheap but systematically miss those dynamics, so naïvely substituting them is unsafe.

We characterize the discrepancy between low- and high-fidelity shock models under impulsive loading, then learn a residual correction with Deep Operator Networks (DeepONets) and recurrent architectures (RNN, LSTM, GRU). Combining transfer learning with residual learning, the resulting multi-fidelity surrogate inherits the speed of the low-fidelity model while recovering the accuracy of the expensive one. The work was conducted under an Office of Naval Research technical report.

Denoising wireless MEMS sensors for infrastructure monitoring

• Wireless sensing
• MEMS
• Generative denoising

Wireless MEMS accelerometers are the workhorse of low-cost infrastructure monitoring, but their low-frequency noise floor obscures the modal signatures and subtle defect signals that engineers actually need to detect.

We calibrate MEMS accelerometers against reference sensors, then train a Generative Adversarial Network to denoise the low-frequency response, comparing against empirical mode decomposition and wavelet baselines. A LabVIEW-based host environment supports durability testing (temperature, salt-spray) and field deployment, including synchronized hardware on tunnel diagnosis vehicles for hidden-defect inspection.

Adaptive vibration control of human-loaded footbridges

• Reinforcement learning
• TD3
• Semi-active TMD

Slender pedestrian footbridges are vulnerable to human-induced vibrations that violate serviceability limits. Conventional passive tuned mass dampers cannot adapt to crowd density, gait variability, or human–structure feedback, leaving residual response under realistic loading.

We build a coupled model of bridge, pedestrian, and semi-active tuned mass damper (STMD), then train a TD3 agent to control the STMD under both periodic and stochastic pedestrian loading. The study reports the influence of key hyperparameters (learning rate, discount factor) on closed-loop performance and proposes a general scheme for RL-based STMD control of pedestrian bridges.

Real-time hybrid simulation under wave and earthquake loading

• Real-time hybrid simulation
• DDPG
• Feedforward compensation

Real-time hybrid simulation (RTHS) of structures under wave and earthquake loading is limited by servo-hydraulic time delay and tracking error, especially in worst-case loading where delay-induced phase lag can destabilize the test. Model-based controllers compensate only partially and depend on accurate plant identification.

We build a Simulink digital twin of an underwater shaking-table system and validate it against physical experiments. A DDPG controller trained against this twin reduces worst-case time delay by 6.54% and tracking error by 7.52% relative to a model-based controller (and by 123.48% / 89.95% relative to no compensation). Adding feedforward compensation to DDPG yields 3.28% maximum perturbation versus 5.88% (PI) and 10.57% (FF only) on a benchmark RTHS problem.

For the underlying papers and links, see the publications page.