Kyeong Sun Kim

How does heterogeneous evidence become trustworthy decisions?

A modern SHM decision — is this offshore foundation healthy? is this bridge safe to keep open? — fuses evidence that arrives in radically different forms: continuous sensor streams, discrete inspection campaigns, expensive engineering simulations, expert judgment, AI-agent recommendations. Each source has its own reliability, cost, time-to-result, and failure mode.

Classical Bayesian fusion treats the likelihoods as fixed. Classical value-of-information is unimodal. Classical attribution ignores the fusion layer entirely. My research program, Trust-weighted Multi-modal Evidence Fusion (TMEF), formalises the missing pieces through four research questions and validates them on structural health monitoring data I collected across centrifuge, simulation, and 32 months of field deployment.

The four research questions I work on:

1. Unified representation. Can we embed vibration, strain, CPT, FEM, and inspection text into one physics-constrained latent manifold without losing modality-specific structure?
2. Trust-weighted fusion. Can SHM Bayesian inference accommodate dynamic, consequence-coupled trust hyperparameters that respond to disagreement and envelope violations?
3. Cross-modal VoI. Can we rank a 10-minute spectral analysis, a 1-week bathymetry, and a 2-week CPT on one modality-invariant cost-per-information scale?
4. Attribution & accountability. When an AI-mediated SHM decision is wrong, can we assign credit and blame across sensor, model, AI, and human in a way that calibrates trust?

The program rests on established foundations — Dempster-Shafer, Jøsang subjective logic, Ades-Sutton evidence synthesis, Straub VoI, Avci SHM fusion. Full program, creative solution approaches, and target venues are on the research page.

What I’ve built

• PhD dissertation (SNU, defense 2026-09) — an integrated numerical and digital-twin framework for scour assessment of offshore wind turbine foundations. 22 centrifuge tests at the KAIST 70g facility, 1,794 Monte Carlo simulations, 32 months of field monitoring on a 4.2 MW turbine in Gunsan, 22,617 analysis windows with zero false alarms. Ch 7’s three-channel Bayesian fusion delivered 67 % posterior σ reduction and 40–70 % realized-cost reduction vs calendar inspection.
• Op³ — open-source Python framework for offshore wind turbine foundation analysis. pip install op3-framework · 140 tests · 39 V&V benchmarks · Zenodo DOI.
• K-Fish — 9-agent LLM swarm with Bayesian confidence-weighted fusion and a Calibrator agent that audits its reasoning. A precursor architecture for TMEF’s multi-agent evidence aggregation with source-specific trust.

About me

BS in Civil & Environmental Engineering at UC Berkeley (2016). PhD at Seoul National University under Prof. Sung-Ryul Kim (defense September 2026). Director at Marine Master Technology Co. Engineer at Myungil Jack-Up Ocean, Korea’s sole manufacturer of modular jack-up barges. Best reached by email.