Open to Work

Yilong
Zhou

Physics & Robotics Simulation Engineer

I simulate physics across four orders of magnitude — from quantum electron structure to robot contact dynamics. My edge isn't the tools. It's knowing when the output is wrong and why.

Multi-scale Sim ML Potentials FEA / FSI Contact Mechanics Soft Robotics Sim-to-Real
PhD · Duke University · Postdoc · LBNL · LLNL
✉ yilongz393@gmail.com View Projects ↓ Resume ↗
4
Sim Scales
100×
MD Speedup
3×
NSF HPC Allocations
17+
Publications
Yilong Zhou
// about

Who I Am

I'm a postdoctoral scholar at Lawrence Berkeley National Laboratory, working at the intersection of computational physics and autonomous systems. My research spans four simulation scales — quantum DFT, molecular dynamics, continuum FEA, and rigid-body robotics — with a focus on knowing what physics breaks down at each scale transition.

I completed my PhD in Mechanical Engineering & Materials Science at Duke University, where I developed custom C++ solvers and deployed large-scale simulations on NSF HPC clusters. I'm now transitioning into industry roles in physical simulation and robotics AI.

My competitive advantage isn't tool proficiency — it's physical intuition: the ability to judge whether a simulation output is physically correct, where a model will fail, and why a policy trained in simulation won't transfer to hardware.

Current Position Postdoctoral scholar · LBNL
PhD Mechanical Engineering & Materials Science · Duke University
Location Bay Area, CA
Target Roles Physics / Robotics Simulation Engineer · Physical AI Researcher
Compute Resources 3× NSF XSEDE · LBNL Savio HPC · Dual RTX 4090
// experience

Where I've Worked

Jan 2025 –
Present
Lawrence Berkeley National Laboratory
Postdoctoral Scholar · Materials Science Division
GPU-accelerated physics simulation pipelines for constrained multi-body dynamical systems on Linux HPC clusters.
  • Built GPU-accelerated simulation pipelines for constrained multi-body dynamical systems on LBNL Savio HPC; validated fidelity through energy-based stability analysis and trajectory convergence metrics
  • Developed automated pipelines coupling simulation runs with downstream analysis, enabling rapid parameter sweeps — directly applicable to environment tuning, reward shaping, and domain randomization in RL training
  • Actively building MuJoCo contact physics models and PPO-based RL policies (Stable-Baselines3) for sim-to-real manipulation tasks — see GitHub →
Jun 2022 –
Nov 2024
Lawrence Livermore National Laboratory
Postdoctoral Researcher · Materials Science Division
ML-augmented physics simulation bridging quantum-scale ab initio calculations and large-scale molecular dynamics.
  • Developed DeePMD neural-network force field framework achieving 100×+ simulation speedup over DFT while preserving physical accuracy; methodology directly analogous to learned dynamics models and sim-to-real transfer in robotics. Published in Physical Chemistry Chemical Physics (2025, First & Corresponding Author)
  • Validated surrogate model accuracy against reference DFT benchmarks across thermodynamic conditions, developing rigorous fidelity assessment protocols — the same physical intuition required to evaluate whether a contact-physics or actuator model is trustworthy in robotics simulation
Aug 2017 –
May 2022
Duke University
Graduate Research Assistant · Arya Lab
Multiscale simulation frameworks linking molecular dynamics to continuum mechanics. Three NSF XSEDE HPC allocation proposals awarded.
  • Designed and validated 3D FSI simulations in COMSOL Multiphysics for soft robotic actuators: modeled pneumatically driven balloon actuators, characterized nonlinear bending dynamics, and used simulation outputs to inform actuation control strategies. Published in Advanced Intelligent Systems (2021) — FSI-Soft-Robotic-Actuator-COMSOL →
  • Pioneered a global Monte Carlo optimization algorithm implemented in C++ from scratch, solving combinatorial search over complex multi-modal energy landscapes — directly applicable to motion planning and configuration-space search in robotics. Published in Nature Communications (2022, Featured Article) — basin-hopping-global-optimizer →
  • Developed ML-based many-body potential for nanoparticle self-assembly using permutationally invariant polynomials fitted to CG MD potentials of mean force — 3000× faster than explicit simulation while reproducing all structural phases. Published in npj Computational Materials (2023) — many-body-nanoparticle-potential →
  • Owned the full simulation-to-validation pipeline: collaborated with experimentalists to benchmark model predictions against physical measurements, building rigorous fidelity verification habits that transfer directly to sim-to-real gap assessment
Sep 2013 –
Dec 2016
Clemson University
Graduate Research Assistant
Derived reduced-order multiphysics models for electrokinetic transport in PDMS microchannels — 4 publications across Electrophoresis, Microfluidics and Nanofluidics, Scientific Reports, and Physics of Fluids.
  • Derived 2D depth-averaged governing equations from second-order asymptotic analysis of full 3D transport (electric, thermal, flow, species, and induced-charge fields) — rigorous reduction that retains top/bottom wall physics, correcting 2–3× threshold errors and wrong wave inclination predictions of naive 2D models; implemented across four COMSOL multiphysics problems (joint simulation work with co-authors). multiphysics-simulation →
  • Performed all experimental work for electrokinetic instability study: PDMS microchannel fabrication by soft lithography, ferrofluid preparation and conductivity characterization across three concentrations, CCD microscopy threshold measurement across four channel depths — directly validated model predictions
  • Implemented coupled COMSOL modules (Electric Currents, Laminar Flow, Heat Transfer in Fluids/Solids, Transport of Diluted Species) with custom depth-averaged wall correction terms added via Force and Reaction features; mesh refinement strategies for multi-scale channel geometries
// selected projects

Projects

Applied simulation engineering across contact physics, reduced-order multiphysics modeling, ML potential deployment, digital twin validation, and RL policy transfer.

Add
mujoco-demo.mp4
MuJoCo Robotics RL / PPO C++
MuJoCo Contact Physics for Sim-to-Real Manipulation
Physically accurate contact model development for rigid and compliant body interaction, with RL policy training targeting sim-to-real transfer.
  • Designed MuJoCo XML models with calibrated contact parameters (stiffness, damping, friction) to match real hardware behavior
  • Implemented PPO-based locomotion/manipulation policies via Stable-Baselines3; identified sim-to-real failure modes from contact model underrepresentation
  • Diagnosed distribution shift between simulated and real contact dynamics using ensemble disagreement metrics — directly analogous to ML potential deployment failures in MD
  • Training on dual RTX 4090 workstation; extending to Isaac Sim for GPU-parallelized rollouts
↗ Ongoing · GitHub repo in progress
GitHub →
ROS2 Turtlesim Demo
ROS2 C++ Jazzy
ROS2 Turtlesim: Catch Them All (C++)
Multi-node ROS2 system where a master turtle autonomously hunts and catches randomly spawned turtles using a proportional controller.
  • 3-node architecture: turtle_spawner, turtle_controller, turtlesim_node communicating via topics and services
  • P controller with angle wrapping for smooth pursuit; closest-target selection via runtime parameter
  • Custom interfaces: Turtle.msg, TurtleArray.msg, CatchTurtle.srv; fully async service calls
  • Planned extension: dynamic target tracking with moving turtles and predictive pursuit control
↗ Coursework final project · C++ implementation · ROS2 Jazzy
GitHub → Certificate ↗
BHMC Optimizer Demo
C++ Global Optimization Monte Carlo HPC
Basin-Hopping Global Optimizer
Custom C++ global optimizer using Basin-Hopping Monte Carlo for ground-state structure discovery — published in Nature Communications (2022, Featured Article).
  • Implemented BHMC from scratch in C++ — 6 custom MC moves designed for quasi-2D interfacial geometries; conjugate gradient local minimization with Wolfe line search
  • Ultrafast shape recognition (USR) for duplicate detection; periodic restart from global best — prevents stagnation in deep local minima
  • Discovered novel A3B5- and A4B6-type 2D superlattices never previously reported; predictions validated against experiment
  • Algorithm maps directly to robotics: energy landscape → cost landscape; basin escape → local minima avoidance in motion planning
↗ Nature Communications 2022 · Featured Article · Open-source C++
Paper ↗ GitHub →
COMSOL FSI Simulation
DraBot wing actuation
COMSOL FSI / Digital Twin Hyperelastic Soft Robotics
FSI Digital Twin for Soft Robotic Actuators
Fully coupled fluid-structure interaction simulation of pneumatic balloon actuators in a soft robotic platform — Mooney-Rivlin/Ogden hyperelastic models, ALE moving mesh, validated against experiment across three channel geometries.
  • Built COMSOL FSI model coupling Laminar Flow + Solid Mechanics modules with ALE moving mesh; second Piola-Kirchhoff framework for large-deformation bending above 80°
  • Fit 2-term Ogden constants from tensile measurements (R²=0.995) for Ecoflex 00-30 — selected over Mooney-Rivlin on physical grounds for very-compliant silicone at large stretch
  • Validated displacement angle vs. injected volume across 3 channel geometries without geometry-specific tuning; max deviation <8%
  • Simulation drove design: parametric geometry study selected 7.1 mm channel for all robot experiments
↗ Published · Adv. Intell. Syst. 2021 · COMSOL FSI by Y. Zhou (sole contributor)
Paper ↗ GitHub →
EXPERIMENT 175.0 V/cm
Experiment at 175 V/cm
periodic waves · inclined upstream ← ✓
DEPTH-AVERAGED 202.1 V/cm
Depth-averaged at 202 V/cm
periodic waves · inclined upstream ← ✓
REGULAR 2D 60.4 V/cm
Regular 2D at 60 V/cm
chaotic at wrong threshold · inclined downstream → ✗
COMSOL Electrokinetics Depth-Averaging FEA Asymptotic Analysis
Electrokinetic Multiphysics — Depth-Averaged Simulation
Derived a 2D depth-averaged model from asymptotic analysis of 3D coupled transport — corrected two simultaneous failures of the naive 2D approach: wrong instability threshold and wrong wave inclination direction.
  • Derived depth-averaged governing equations for five coupled fields (electric, thermal, flow, species, ICEO) via second-order asymptotic expansion in δ = d/H — retains top/bottom wall physics entirely absent from regular 2D models; joint simulation work with co-authors
  • Regular 2D model failed two ways independently: 2–3× threshold under-prediction and instability waves inclined downstream (→) vs experiment (←); single wall correction term fixed both simultaneously
  • Validated across four channel depths (32–100 µm) and three ferrofluid concentrations; 6–15% error vs 49–80% for regular 2D in shallow channels (d/W < 0.3)
  • Performed experimental work: PDMS microchannel fabrication by soft lithography, ferrofluid preparation, threshold measurement by CCD microscopy
↗ 4 publications · 10 min/run vs >4 hr for equivalent 3D model
GitHub → Sci. Rep. ↗ Phys. Fluids ↗ Electrophoresis ↗ Microfluid. ↗
CG GROUND TRUTH ~15,500 CPU-hrs
CG explicit ground truth — 1D strings
1D strings ✓
MANY-BODY POTENTIAL ~5 CPU-hrs
Many-body potential — 1D strings
1D strings ✓ · 3000× faster
ML Potential LAMMPS PIPs Python MBX
Many-Body Nanoparticle Potential via Machine Learning
Physics-informed ML surrogate for nanoparticle self-assembly — 3 orders of magnitude faster than explicit simulation while preserving the physics that drives anisotropic assembly.
  • Computed two- and three-body potentials of mean force (PMFs) from ~22,000-atom CG MD via blue moon ensemble; integrated constraint forces to build PIP training data
  • Fitted permutationally invariant polynomials (PIPs) via ridge regression with energy-weighted residuals; adapted MB-Fit/MBX framework (Paesani group) from molecular to mesoscale NP systems
  • Key validation: two-body-only model predicts wrong structure (closed triangles); three-body term is what makes 1D strings stable — physically necessary, not a fitting artifact
  • Deployed via LAMMPS pair_style mbx; 3000× speedup enabled discovery of gel, network, and cluster phases inaccessible at explicit CG timescales
  • Validated assembly pathways — not just final structures: many-body potential reproduces the correct route to string formation (dimer → linear extension), while two-body collapses to wrong closed-triangle metastable state at both the energy landscape and dynamics levels
↗ npj Computational Materials 2023
Paper ↗ GitHub →
// core differentiator

Cross-Scale Simulation Stack

Most engineers operate at one scale. I build pipelines that span all four — and understand the physics that breaks down at every handoff.

01 / Quantum
DFT
Å – nm
VASP PBE / PAW AIMD
Electronic structure, forces, bonding environments. Output feeds ML potential training.
→
02 / Atomistic
MD
nm – μm
LAMMPS HOOMD-blue DeePMD GPU MD
ML potentials enable ab initio accuracy at classical cost. Ensemble disagreement detects distribution shift.
→
03 / Continuum
FEA / FSI
μm – mm
COMSOL FSI Thermal-Mech Digital Twin
Continuum mechanics and coupled multiphysics. Digital twins validated against experiment.
→
04 / Robotics
Contact
mm – m
MuJoCo Isaac Sim ROS2 PPO / RL
Rigid and compliant body dynamics. Sim-to-real gap diagnosis from physics first principles.
// technical skills

Stack

Quantum · DFT
VASP DFT / AIMD PBE · PAW vdW corrections Electronic structure
Atomistic · MD
LAMMPS HOOMD-blue DeePMD-kit ML potentials GPU-MD
Continuum · FEA
COMSOL FSI Electrokinetics Thermal-Mech Digital Twin
Robotics Sim
MuJoCo Isaac Sim ROS2 PPO / RL Stable-Baselines3
Programming
C++ · MPI Python Custom solvers HPC pipelines 3× NSF XSEDE
// research record

Publications

17+ papers · H-index 14+ · Full list on Google Scholar ↗

Nature Comms
Discovery of Two-Dimensional Binary Nanoparticle Superlattices Using Global Monte Carlo Optimization
Nature Communications · 2022 · Featured Article
C++ global optimizer — directly maps to motion planning over cost landscapes
npj Comp Mat
Many-body Potential for Simulating the Self-Assembly of Polymer-Grafted Nanoparticles
npj Computational Materials · 2023
ML surrogate modeling for complex many-body dynamics
Adv. Intel. Sys.
Microengineered Materials with Self-Healing Features for Soft Robotics
Advanced Intelligent Systems · 2021
COMSOL FSI digital twin for balloon actuator — fluid-structure coupling, large-deformation mechanics, validated against experiment
Phys. Fluids
Induced Charge Effects on Electrokinetic Entry Flow
Physics of Fluids · 2017
Physics-based simulation of fluid flow in microfluidic devices — COMSOL model validated against experiment; published in top fluid mechanics journal
Nature Comms
Self-assembly of nanocrystal checkerboard patterns via non-specific interactions
Nature Communications · 2024 · Co-first Author · Featured Article
Simulation-experiment co-validation · covered by Phys.org, EurekAlert
Langmuir
From Frustration to Order: Role of Fluid-Fluid Interfaces in Precision Assembly of Nanoparticles
Langmuir · 2024 · Invited Perspective by Editor-in-Chief
Invited by editor-in-chief — field recognition of multi-scale simulation expertise
// get in touch

Let's Build Together

Physics / Robotics Simulation Engineer · Robotics AI.
Open to conversations about simulation pipelines, sim-to-real transfer, and autonomous systems.

✉ Email ↗ LinkedIn ↗ GitHub ↗ Google Scholar