Home / Work

Skills, tools & results

I generate quantum reference data, train models of atomic interactions, and run simulations on supercomputers. I check the predictions against reference calculations and available experiments.

Results

Simulation and validation results

25
dopants

Benchmarked a universal ML potential, open source

Compared Meta’s UMA with DFT for 25 MoS₂ dopants (formation-energy MAE 0.374 eV for UMA small), then simulated heating and cooling. With UMA small, relaxing the structures and calculating formation energies for all 25 dopants took 0.53–0.97 hours, depending on the GPU. One ML potential, 25 dopants: where it works and where it fails →

76 → 9.2
meV/atom

Fine-tuned a foundation model with LoRA

Energy error of a MACE model for MoS₂ with nitrogen dropped eightfold after low-rank (LoRA) fine-tuning on automatically generated DFT data. Paper in preparation. Fine-tuning cuts a foundation model's error eightfold for MoS₂ + N₂ →

0.034
R-factor

Matched X-ray diffraction of a molten salt

Repaired an unphysical Fe–Fe clustering artifact in a universal MACE potential for molten LiCl–FeCl₂, then matched high-energy X-ray diffraction. Under review. Fixing a foundation model's iron clustering in molten salt →

220 K
transition

Identified a low-temperature drop in MoS₂ lubricant performance

Compiled friction, wear and wear-life data from 21 references. In vacuum or inert gas, reported performance deteriorated below about 220 K. A companion simulation study tested why: smaller cold-formed flakes adsorbed more water and aligned less readily with sliding. MoS₂ lubricant performance across temperatures →

What I deliver

An end-to-end ML potential pipeline

1 · Data

Automated DFT frame generation across compositions and temperatures. Active learning adds only the configurations the model is unsure about.

VASPQuantum ESPRESSOASEActive learning

2 · Models

Train from scratch or fine-tune foundation models (MACE, UMA, eSEN). LoRA and multi-head replay adapt a model to new chemistry while retaining its broader training.

MACEUMAeSENDeepMDPyTorch

3 · Simulation & validation

GPU MD of thousands of atoms on leadership-class machines, checked against DFT, AIMD, X-ray diffraction, and tribometer data.

LAMMPSAuroraPolarisSlurm / PBS
Toolbox

Skills matrix

Machine-learning potentials

MACEUMA (fairchem)eSENDeepMD-kitLoRA fine-tuningMulti-head replayActive learningUncertainty estimation

Simulation

LAMMPSASEVASPQuantum ESPRESSOReaxFFEAM / MEAMMetadynamicsFree-energy profiles

Data & code

PythonPyTorchNumPySciPypandasscikit-learnMatplotlibMATLAB

HPC & workflow

ALCF AuroraALCF PolarisSlurmPBSMPIGPU scalingHigh-throughput DFTAgentic workflows

Analysis & visualization

OVITOVMDKirkwood–Buff theoryRDF / structure factorMSD / diffusionParity & error analysis
Experience

Where I have done it

Fall 2026 – present

Guest Facility User

Argonne National Laboratory · Center for Nanoscale Materials
  • Train and fine-tune MACE, MACE-RSGA, and eSEN models for platinum–water interfaces.
  • Compare reaction barriers and solvent structure against DFT and ab initio MD.
  • Use active learning to add the atomic environments the models are missing.
MACEeSENPt–waterAIMDActive learning
Mar 2026 – Aug 2026

Research Aide Technical – PhD

Argonne National Laboratory · Computational Science Division
  • Fine-tuned universal MACE potentials for molten LiCl–FeCl₂ with multi-head replay and LoRA. This removed an unphysical Fe–Fe clustering artifact inherited from solid-state training data.
  • Validated liquid structure against high-energy X-ray diffraction (mean R-factor 0.034) across composition and temperature.
  • Automated DFT frame generation for labeled training data across molten-salt compositions on ALCF systems.
  • Built an uncertainty-biased active-learning loop: bias forces push MD toward uncertain configurations, which are DFT-labeled and added to training.
  • Computed excess thermodynamic properties and the LiCl–KCl phase diagram with Kirkwood–Buff theory on ML-driven MD.
  • Trained and benchmarked MACE foundation models on the Open Catalyst 2025 (OC25) dataset at Aurora scale. Built agentic workflows that set up data generation and fine-tuning from plain-language instructions.
MACELoRAAuroraPolarisActive learningKirkwood–Buff
Mar 2025

Visiting Scholar

NASA Jet Propulsion Laboratory
  • Ran tribometry of MoS₂ dry-film coatings across temperatures.
  • Fed the observations into the low-temperature wear simulations.
TribometryMoS₂Space lubricants
Jan 2023 – present

Graduate Student Researcher

Martini Research Group · University of California, Merced
  • Fine-tuned a MACE model for MoS₂–N₂ with LoRA and automated DFT data. Validation energy RMSE fell from 76.0 to 9.2 meV/atom.
  • Benchmarked Meta’s universal UMA potential against DFT for 25 MoS₂ dopants and published the workflow as open source (JCTC 2026).
  • Trained DeepMD potentials for ions at solid–liquid electrochemical interfaces.
  • Co-originated, and helped benchmark the ML potential for, a study of ion adsorption on gold (JCP 2026).
  • Ran fracture, water-adsorption, and sliding simulations that explain cold-temperature failure of MoS₂ lubricants (Friction 2025).
UMAMACEDeepMDReaxFFLAMMPSMetadynamics
Jun 2022 – Dec 2022

Research Assistant

Motalab Lab · BUET
  • Related cooling rate and microstructure to void nucleation and cyclic response.
  • Mentored undergraduates in simulation and analysis. The work led to peer-reviewed papers on solder, CdTe nanowires, and printed Inconel.
LAMMPSEAMMEAMMentoring
Let's talk

Open to Summer 2027 research internships and fellowships

Atomistic modeling, ML interatomic potentials, and HPC workflow roles in industry R&D and national laboratories. Enrolled at a U.S. university through Fall 2027.