AI-Driven Design & Digital Twin

Physics-informed AI surrogates that turn days of simulation into seconds — for turbine, automotive, and semiconductor design.

Why it matters

Finite element analysis can predict how a component will deform, crack, or wear out — but a single high-fidelity run may take hours, and a design optimization needs thousands of them. Machine learning can be fast, but a model that ignores the underlying mechanics is not to be trusted near the edge of its training data. The opportunity lies in combining the two: surrogate models that learn from simulation while respecting the physics, so that engineers can evaluate and improve designs in real time. Our lab brings two decades of computational mechanics — from atomistic simulation to continuum fracture — to this problem.

What we do

FEA result versus AI prediction of stress field on a mechanical component

Figure 1. Finite element stress field (left) and prediction by a physics-informed graph neural network (right).

Physics-informed graph neural networks for structural design. We build surrogate models that operate directly on unstructured finite element meshes: an attention-based 3D graph neural network predicts displacement and stress fields node by node, and its training loss is derived from the total potential energy of the structure so that predictions remain mechanically consistent. Coupled with multi-objective optimization and automated CAE pipelines, the tool lets a designer explore geometry, cooling, and material choices for high-temperature components in seconds rather than days. Our current focus is the structural design and life prediction of gas-turbine components.

Phase-field fracture and digital twins of coated components. Where a crack will start and how it will grow is exactly the question surrogate models must eventually answer. We implement phase-field fracture models in ABAQUS to simulate spallation of thermal barrier coatings and interface cracking under thermal cycling, generating the high-fidelity data that AI models learn from and validating their predictions.

Schematic of the nanoscale field projection method

Figure 2. Field projection method: cohesive-zone laws are extracted from molecular dynamics for use in continuum-scale simulation.

Scale bridging from atoms to continuum. Long before "surrogate model" became a common phrase, we were compressing atomistic physics into continuum laws. Our nanoscale field projection method extracts cohesive tractions, separations, and surface stresses from molecular dynamics data near a crack tip, yielding mixed-mode cohesive-zone laws for grain boundaries in metals and graphene that can be used directly in finite element models. Related work includes extended JKR theories of adhesive contact for coated bodies and dislocation nucleation under nano-asperity contact.

Mesoscale and coarse-grained simulation of polymers. Dissipative particle dynamics and coarse-grained molecular dynamics let us study tribology of polymer-brushed surfaces and polymer pyrolysis at length and time scales inaccessible to atomistic methods — supporting materials design in our fiber and actuator programs.

Key capabilities

Physics-informed graph neural networks on unstructured FE meshes (PyTorch Geometric)
Automated CAE pipelines: ANSYS/PyAnsys, ABAQUS with user subroutines (UMAT, UEL)
Multi-objective evolutionary optimization and active-learning sampling
Phase-field fracture modeling of interfaces and coatings
Molecular dynamics (LAMMPS), dissipative particle dynamics, coarse-grained MD
Nanoscale field projection and cohesive-zone law extraction

Selected publications

V. P. Nguyen, I. Jeon, S. Yang*, and S. T. Choi*, Mesoscale simulation of polymer pyrolysis by coarse-grained molecular dynamics: A parametric study, ACS Applied Materials & Interfaces, 2023. [DOI]
S. T. Choi*, N. T. Mai, and V. P. Nguyen, Dislocation nucleation and segregation under adhesive contact of a nano-asperity coating on a crystalline solid, European Journal of Mechanics A/Solids, 2021. [DOI]
V. P. Nguyen, N. T. Mai, and S. T. Choi*, Atomic mixed-mode cohesive-zone dual constitutive laws of impurity-embrittled grain boundaries in polycrystalline solids via nanoscale field projection method, Journal of the Mechanics and Physics of Solids, 2021. [DOI]
V. P. Nguyen and S. T. Choi*, Extended JKR theory on adhesive contact of coated spheres, Acta Mechanica, 2019. [DOI]
V. P. Nguyen, P. Q. Phi, and S. T. Choi*, Tribological behaviors of grafted-nanoparticle on polymer-brushed walls: A dissipative particle dynamics study, ACS Applied Materials & Interfaces, 2019. [DOI]
N. T. Mai, P. Q. Phi, V. P. Nguyen, and S. T. Choi*, Atomic-scale mode separation for mixed-mode intergranular fracture in polycrystalline metals, Theoretical and Applied Fracture Mechanics, 2018. [DOI]
N. T. Mai and S. T. Choi*, Atomic-scale mutual integrals for mixed-mode fracture: Abnormal fracture toughness of grain boundaries in graphene, International Journal of Solids and Structures, 2018. [DOI]
S. T. Choi and K.-S. Kim*, Nanoscale planar field projections of atomistic decohesion and slip in crystalline solids, Part I: A crack-tip cohesive zone, Philosophical Magazine, 2007. [DOI]

Full list: Publications

Industry & careers

This track connects to power generation and aerospace propulsion, automotive CAE groups, semiconductor packaging reliability, and CAE and engineering-software companies. Students learn to speak both languages — finite element mechanics and machine learning — and to build the automated pipelines that connect them, a profile in high demand across Korean heavy industry and increasingly in software. Typical career paths include CAE / simulation engineer, machine-learning engineer for engineering design, digital-twin developer, and computational materials scientist.

FMAM Lab

AI-Driven Design & Digital Twin