Personal Homepage

Hyeongmin Kim

ohmnum, the chick mascot

Hi, I’m Hyeongmin — a Staff Engineer at Samsung Research.

I research how to teach AI models what we know, and how we want them to act.

Email CV Google Scholar

GitHub LinkedIn ORCID

GCP ML Engineer

NEON VOID Browser space shooter — a side project ↓ v1.1.0

About

A quick intro.

Before: Domain Knowledge-Informed Neural Networks. Building physics and expert knowledge into models for machine fault diagnosis (Ph.D., Seoul National University).

Now: Data-based BrainWash. Shaping how LLMs behave through the data they learn from.

Skills

Toolbox

Domains

PHM Fault Diagnosis Motors / PMSM Rotating Machinery Thermal Power Plants Engineering Design

Methods

Deep Learning Signal Processing Domain-informed NN (DINN) Anomaly Detection Domain Adaptation Self-/Semi-supervised Reinforcement Learning Inverse Modeling Uncertainty Estimation

AI Systems

LLM Behavior Correction Post-training Training Data Selection LLM / RAG Knowledge Graphs LLM Agents Multi-modal LLM

Tools

Python PyTorch

Research

What I Work On

Research details All publications

There are three ways to get knowledge into a model: through the data it learns from, through its structure, and through what it can look up when it answers. The first three areas below follow that order.

01 · Current focus

Data-based BrainWash: LLM Behavior Correction

Correcting specific unwanted behaviors of deployed LLMs at low cost through training techniques and careful data selection and composition — while preserving safety behavior and general capability.

Training techniques for targeted behavior correctionresearch
Data selection & composition for correction trainingresearch
Side-effect evaluation — safety refusal, factuality, and general ability preservationresearch

Domain Knowledge Informed Neural Networks

Networks guided by physics, signal analysis, and expert priors — for generalization, physical consistency, and interpretability.

DINN review — survey and taxonomy of domain-informed PHM methodsACME ’26 (accepted)
Physically-consistent models (PINN-style losses)theme
Interpretable inputs (2D time–frequency maps)theme

Knowledge Graphs & Retrieval

Structuring domain knowledge into graphs so LLMs can find and use it when they answer.

OrcheRAG: multi-constraint product search with GraphRAGproject
Knowledge graph + RAG for PHMresearch

Prognostics & Health Management

Fault diagnosis, anomaly detection, and predictive maintenance for rotating machines and thermal power plants.

SCOC — preprocessing for motor fault diagnosis under variable conditions Measurement ’23
MPARN — multi-scale path attention residual network JCDE ’23
Opt-TCAE — boiler-tube leakage detection in thermal plants ESWA ’23
Center-margin, uncertainty-aware diagnosis of unseen faults JCDE ’26

AI-driven Design

Generative design, performance prediction, and parameter / topology optimization via inverse modeling and reinforcement learning.

Deep learning-based die design to compensate springback in bipolar-plate stamping (w/ Fraunhofer IPT) Measurement ’26
Design evaluation with PINNs & explainable AItheme
RL-based design-parameter optimizationtheme

Experience

Career & Education

Full CV

Work

Staff EngineerSamsung Research Now
Postdoctoral ResearcherHAI Lab, Seoul National University 2024
Part-time InternOnepredict Inc. 2023
Visiting ResearcherFraunhofer IPT, Aachen 2022
Undergraduate InternBiorobotics Lab, SNU 2017

Education

Ph.D., Mechanical EngineeringSeoul National University · Adviser: B. D. Youn 2019–24
B.S., Mechanical & Aerospace Eng.Seoul National University 2015–19

Certification

Professional Machine Learning Engineer Google Cloud · Issued Jun 2026 · Expires Jun 2028 Verify ↗

Research Output

Publications, Patents & Awards

Google Scholar

227 citations · h-index 7 · i10-index 7 (Google Scholar, September 2026)

Publications

14 journal articles and 15 peer-reviewed conference papers in PHM, fault diagnosis, and domain-informed AI.

View publications → 8

Patents

Registered and filed Korean patents on motor fault diagnosis, boiler-tube leakage detection, and anomaly detection.

View patents → 10

Awards

Doctoral dissertation, first-author paper, and team competition awards from KSME, PHMAP, and IEEE PHM.

View awards →

Activities

Talks & Presentations

May 2024

Window Size and Sampling Rate Selection for Cost-optimal Deep Learning-based Fault Diagnosis

ACSMO 2024 · Zhengzhou, China
Aug 2023

Convolutional Auto-Encoder-Based Boiler Tube Leakage Detection in a Thermal Power Plant

ASME IDETC-CIE 2023 · Boston, USA
Aug 2022

A Multi-scale Residual Network with Attention Mechanism for Fault Diagnosis of Rotating Machines

ASME IDETC-CIE 2022 · St. Louis, USA
Aug 2022

Reinforcement Learning-based Optimal Parameter Estimation of Ultra Precision Machine

ICEAS 2022 · Ansan, South Korea
Sep 2021

Anomaly Detection of Multichannel Complex Industrial Systems Using Physically Grouped Convolutional Autoencoder

PHMAP 2021 · Jeju, South Korea

Side Project

Built for fun

Browser Game · WebGL

NEON VOID v1.1.0

Strap the ohmnum chick into a neon fighter and clobber the void’s nastiest space monsters — dodge bullet-hell waves, grab power-ups, chain combos, and smash a boss every five stages. Right in the browser, with a global leaderboard.

▶ Play now Game wiki ↗

Extra

Older notes and study logs.

I keep older lecture notes and study materials in a separate archive so this homepage can stay focused while the earlier material remains available.

Open extra archive

Contact

Get in touch.

HyperAutomation AI Lab · Rm 324-2, Bldg 313
Seoul National University, Seoul 08826

960802@snu.ac.kr Google Scholar ORCID · 0000-0003-2693-4621 LinkedIn

GitHub / ohmnum HAI LAB