M.S. Student · KAIST

Jinseong Han

Researching autonomous design for electric vehicles based on artificial intelligence.

  • Agentic design
  • Multi-physics
  • Electric motors
Portrait of Jinseong Han
01

About

Education

  1. Korea Advanced Institute of Science and Technology (KAIST) M.S. in CCS Graduate School of Mobility · Adviser: Prof. Namwoo Kang Daejeon, South Korea
  2. Dongguk University B.S. in Mechanical Engineering · Adviser: Prof. Soohwan Park Seoul, South Korea

Experience

  1. Graduate Researcher (M.S. Candidate) Smart Design Lab, KAIST Daejeon, South Korea
  2. Student Researcher Electrified Propulsion Lab. Seoul, South Korea

Awards

Paper of Distinction Award Design Automation Conference (DAC)
ASME IDETC/CIE 2026 For “A Multi-Agent System for IPMSM Design Optimization via an FEA-AI Hybrid Approach”

Projects

  • Hyundai Motor Group Development of AI-Based Electromagnetic Analysis Models and Performance Prediction for IPMSM Jun. 2025 – Jun. 2026
  • Korea Atomic Energy Research Institute (KAERI) Development of AI-based Simulation Models for Predicting Three-Dimensional Unsteady Flow Fields Jan. 2025 – May 2026

Patents

  • Method for Building Artificial Intelligence Model for Optimization of Design Shape and Electronic Device and System Implementing Such Method KR Application No. 10-2026-0102200
  • Method for Automatic Optimization of Design Shape Using Multi-Agent and Electronic Device and System Implementing Such Method KR Application No. 10-2026-0102201
02

Research

Framework overview: agents for design problem definition, AI surrogate training, and optimization with an FEA–AI hybrid model
  • Paper of Distinction · IDETC 2026
  • arXiv
  • Live demo

01 Multi-Agent based Parametric Motor Design Optimization

LLM agents define the optimization problem, repair the design space, train an uncertainty-aware surrogate and run an FEA–AI hybrid optimizer that calls FEA only where the AI is unsure — end to end.

28 → 84%
Geometry feasibility
−44%
Iron loss
−52–55%
Computational cost
Figures & details Live demo arXiv
Design, generation and evaluation agents for non-parametric motor design 02 Multi-Agent based Non-parametric Motor Design Optimization Agents generate and edit rotor images beyond fixed parameters, then turn them into STEP geometry for FEA–AI evaluation. Generative design · LLM agents Ground-truth and predicted velocity fields around a multi-cylinder array at three time instants 03 Spatio-Temporal Multi-cylinder Flow Prediction A coordinate-based network with positional time encoding predicts unsteady flow fields around multi-cylinder arrays. CFD surrogate · Neural fields A 3D bracket aligned to CNC axes and its three binary projections 04 Manufacturing-Aware 3D Geometry Post-Processing for 3-Axis CNC Milling Depth-map projections turn free-form 3D geometry into CAD models that a 3-axis CNC mill can machine. Manufacturability · CAD reconstruction Motor geometry mapped to a vibration response surface with a conditional VAE 05 Electromagnetic–Vibration Multiphysics Prediction of IPMSM A conditional VAE with 2D→3D parameter transfer predicts motor vibration response from design variables. Multi-physics · Transfer learning
03

Publications

Google Scholar

International journal

  1. 2026

    A LLM-Based Multi-Agent System for Motor Design Optimization Using an Uncertainty-Aware FEA-AI Hybrid Model

    Han, J., Yang, S.*, and Kang, N.**

    arXiv:2606.09037 Under review arXiv Demo

International conference

  1. 2026

    A Multi-Agent System for IPMSM Design Optimization via an FEA-AI Hybrid Approach

    Han, J., Yang, S., and Kang, N.*

    ASME 2026 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference (IDETC/CIE 2026), DETC2026-186939 Paper of Distinction

Domestic conference

  1. 2026

    A Multi-Agent System-Based Non-Parametric Image Generation for IPMSM Design Optimization

    Han, J., Yang, S., and Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

  2. 2025

    IPMSM Shape Optimization based FEA-AI Hybrid Model using LLM agents

    Han, J., Yang, S., and Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

  3. 2025

    Time-Series Flow Field Prediction of Multi-Cylinder using Spatio-Temporal Coordinate-based Neural Network

    Han, J., Kim, S., Kim, S., Yang, S., Cho, B., Song, C., Kang, J., Song, M., and Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

  4. 2026

    Continuous Flow for Discrete Topology Generative Design Editing of IPMSM Rotors via VQ-VFM

    Jeong, L., Han, J., …, Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

  5. 2025

    Manufacturing-Aware 3D Geometry Post-Processing Framework via Depth Maps for 3-Axis CNC Milling

    Kim, S., Han, J., Kim, S., Cho, Y., and Kang, N.*

    Domestic conference

  6. 2025

    Transformer-based Framework for Large-Scale 3D Flow Prediction

    Kim, S., Yang, S., Kim, S., Han, J., and Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

  7. 2025

    Spatio-Temporal Flow Field Prediction of Variable Multi-Cylinder Configurations via DeepONet

    Kim, S., Yang, S., Kim, S., Han, J., Cho, B., Song, C., Kang, J., and Kang, N.*

    The Korean Society of Mechanical Engineers (KSME)

04

Live demo

MotorAI is an interactive console for the multi-agent IPMSM optimization framework.

  1. Design agentTurns a chat into a structured problem: objective, design variables, constraints.
  2. Training agentResamples failed and sparse regions, then trains a UQ-aware deep-ensemble surrogate.
  3. Optimization agentAn LLM sets the switching threshold τ that routes each evaluation to FEA or AI.
Open full demo Source Runs in the browser · KO / EN
hanjinseong.github.io/ipmsm-design

Research 01

Multi-Agent based Parametric Motor Design Optimization

  • Paper of Distinction · IDETC 2026
  • LLM agents
  • FEA–AI hybrid
  • Uncertainty
  • IPMSM

Interior permanent magnet synchronous motor (IPMSM) design requires balancing conflicting objectives and multi-physics constraints, while modern optimization workflows face three bottlenecks: manual problem setup, high finite element analysis (FEA) cost, and unreliable surrogate-based search in sparse or out-of-distribution regions. To address these limitations, we propose an end-to-end automated IPMSM design optimization framework that integrates retrieval-augmented generation (RAG) for structured problem definition with an uncertainty-aware FEA-AI hybrid optimization pipeline.

<50 → 67–80%
answer accuracy on motor-design problems with RAG (Design agent)
28 → 84%
geometry-sampling success after the Training agent repairs the design space
−44%
iron loss (up to) vs. FEA-only search under the same FEA budget
+22.5%
hypervolume in multi-objective optimization, same FEA budget
−52–55%
computation time while retaining 90–92% of FEA-only hypervolume
−5.8%
iron loss with the adaptive uncertainty threshold vs. a fixed one
Three-step framework: design problem definition by the problem-definition and design-sampling agents, AI surrogate training by the FEA and training agents, and GA optimization using an FEA–AI hybrid model
Fig. 1 Agent-based IPMSM design optimization — design problem definition, AI surrogate training, and optimization with the FEA–AI hybrid model.
IPMSM cross-section annotated with design variables x1 to x12 and a table of sampled designs
Fig. 2 IPMSM design space — geometric design variables and sampled design points.

Related publications

  • Han, J., Yang, S., & Kang, N. (2026). A Multi-Agent System for IPMSM Design Optimization via an FEA-AI Hybrid Approach. arXiv preprint arXiv:2606.09037.
  • Han, J., Yang, S., and Kang, N.* (2026). A Multi-Agent System for IPMSM Design Optimization via an FEA-AI Hybrid Approach. ASME IDETC/CIE 2026, DETC2026-186939. Paper of Distinction
  • Han, J., Yang, S., and Kang, N.* (2025). IPMSM Shape Optimization based FEA-AI Hybrid Model using LLM agents. KSME.
Live demo arXiv:2606.09037

Research 02

Multi-Agent based Non-parametric Motor Design Optimization

  • Generative design
  • LLM agents
  • Image → STEP
  • IPMSM

Moving beyond a fixed set of design parameters, this work lets agents explore motor shapes directly as images. A design agent sets up the problem from a reference image, the objective functions and the design space. The generation agent works in three parts: an exploration agent draws new rotor images with a text-to-image model, an exploitation agent turns measured performance into an improvement direction for a fine-tuned model and repairs broken shapes, and a filtering agent keeps the promising candidates using a multi-modal embedding of each image and a description of it. The evaluation agent extracts CAD from the image — RGB to binary to line-and-fillet to STEP — and scores it with the uncertainty-aware FEA–AI hybrid model, calling FEA where the surrogate is unsure.

Pipeline of a design agent, a generation agent made of exploration, exploitation and filtering agents, and an evaluation agent doing CAD extraction and FEA–AI evaluation
Fig. 1 Multi-agent based IPMSM design — design, generation (exploration · exploitation · filtering) and evaluation agents.
Baseline motor geometry and four edited non-parametric rotor images
Fig. 2 Non-parametric image editing — from the baseline geometry to edited 2D rotor images.

Related publication

  • Han, J., Yang, S., and Kang, N.* (2026). A Multi-Agent System-Based Non-Parametric Image Generation for IPMSM Design Optimization. KSME.

Research 03

Spatio-Temporal Multi-cylinder Flow Prediction

  • CFD surrogate
  • Coordinate-based network
  • Positional encoding

A spatio-temporal coordinate-based neural network predicts time-series flow fields around multi-cylinder arrays. Time values are embedded with sinusoidal positional encoding, giving the network a rich representation of time, and the predicted velocity-magnitude fields are compared with the CFD ground truth at several instants of the unsteady flow.

Time values embedded into a positional-encoding time vector using sine and cosine functions
Fig. 1 Positional encoding of time values into a PE time vector.
Ground truth and predicted velocity magnitude at 13.15, 22.03 and 37.45 seconds
Fig. 2 Ground truth vs. predicted velocity magnitude at 13.15 s, 22.03 s and 37.45 s.

Related publications

  • Han, J., Kim, S., Kim, S., Yang, S., Cho, B., Song, C., Kang, J., Song, M., and Kang, N.* (2025). Time-Series Flow Field Prediction of Multi-Cylinder using Spatio-Temporal Coordinate-based Neural Network. KSME.
  • Kim, S., Yang, S., Kim, S., Han, J., Cho, B., Song, C., Kang, J., and Kang, N.* (2025). Spatio-Temporal Flow Field Prediction of Variable Multi-Cylinder Configurations via DeepONet. KSME.

Research 04

Manufacturing-Aware 3D Geometry Post-Processing Framework via Depth Maps for 3-Axis CNC Milling

  • Manufacturability
  • CAD reconstruction
  • 3-axis CNC

A post-processing framework that turns free-form 3D geometry into shapes a 3-axis CNC mill can actually machine, in three steps: (1) project the geometry along the three fixed machining axes and extract binary contour maps; (2) parameterize the 2D outlines with SDF-based contour smoothing, polyline simplification and line–fillet fitting; (3) reconstruct the CAD model with Boolean subtraction and refine its parameters with IoU-based optimization.

Three-step framework: 2D projection contour extraction, 2D outline parameterization, 3D reconstruction and shape optimization
Fig. 1 Framework — from a raw 3D geometry to a machinable CAD model.
3D geometry aligned to the axes of a 3-axis CNC machine and its binary projections along X, Y and Z
Fig. 2 2D–3D optimization — axis alignment for 3-view modeling and the resulting 2D binary projections.

Related publication

  • Kim, S., Han, J., Kim, S., Cho, Y., and Kang, N.* (2025). Manufacturing-Aware 3D Geometry Post-Processing Framework via Depth Maps for 3-Axis CNC Milling.

Research 05

Electromagnetic–Vibration Multiphysics Prediction of IPMSM

  • Multi-physics
  • Conditional VAE
  • 2D → 3D transfer learning

A generative surrogate that maps IPMSM design variables to the motor's electromagnetic–vibration response. A conditional variational autoencoder (CVAE) is first trained on a 2D analysis dataset; its encoder and decoder parameters are then transferred to a second stage that learns 3D analysis results, conditioned on the same design variables.

Motor cross-section mapped to a vibration response surface, and a conditional VAE trained on 2D data with parameter transfer to 3D training
Fig. 1 Electromagnetic–vibration prediction — CVAE with parameter transfer from 2D to 3D training.