From electronic structure to executable workflows.

01 · Workflow systems

ParslBox

ParslBox is a Python toolkit built on Parsl for automated and agentic scientific workflows on HPC systems. I created it to coordinate machine-learned interatomic potential (MLIP) campaigns whose data generation, training, evaluation, molecular dynamics, and analysis stages require different applications and resource layouts.

It combines a local SQLite job database, command-line and Python APIs, CPU/GPU affinity-aware resource management, configurations for Aurora, Polaris, and Sophia, plugins for LAMMPS, VASP, MACE, and ACE, and a Model Context Protocol (MCP) server for agent-driven execution. Current testing reaches 2,000 nodes and 24,000 GPUs.

Repository

02 · Atomistic simulation

Molten-salt MLIPs for CSTEEL

I lead development and application of machine-learned interatomic potentials for FeCl₂, FeCl₃, and FeCl₂-LiCl mixtures within the Center for Steel Electrification by Electrosynthesis (CSTEEL). The workflow uses thousands of PBE and r²SCAN DFT calculations, active learning, MACE foundation models, fine-tuning, and ACE distillation.

MLIP-driven molecular dynamics resolves iron coordination, diffusion, radial distribution functions, structure factors, polymer-chain formation, and liquid connectivity. Comparisons with high-energy X-ray diffraction provide experimental validation and support interpretation of molten electrolyte structure for electrochemical steelmaking.

03 · Graph learning

Redox-potential GNN

This completed project built a high-throughput framework for predicting redox potentials of Fe(II)/Fe(III) complexes. I generated a consistent DFT dataset of roughly 2,300 iron complexes from tmQM and developed the molecular graph construction and PyTorch training pipeline using graph convolutional and graph attention architectures.

The models achieved an error margin of about ±0.2 V, enabling rapid prioritization of ligand and complex families without a new DFT calculation for every candidate. The primary study is published; a follow-up ensemble and uncertainty study is under review.

04 · Thermal transport

AI4NS-Composites

I build scalable workflows and simulation capability for thermal and mechanical properties of binary composites, contributing to an agentic platform that can launch a materials workflow when a requested property is not already available.

For Al-SiC and Cu-SiC systems, I distilled ACE force fields from MACE foundation models and used them to compute thermal conductivity in bulk phases and interfaces. The molecular dynamics simulations reach 150,000 atoms, making interface-scale calculations tractable at a cost impractical for the teacher model.

05 · Vision-language models

Cyclic-voltammetry vision-language modeling

As computational scientist, ML lead, and co-PI, I designed an end-to-end pipeline for extracting electrochemistry knowledge from cyclic-voltammetry figures. It searches open-access literature, detects curves in complex layouts, separates them from axes and labels, reconstructs current-potential data, and formats expert annotations for training.

I led foundation-model selection, data preparation, fine-tuning, evaluation, and iteration. The $50k Argonne Laboratory Directed Research and Development seed project produced a complete proof of concept from published figures to structured data and specialized fine-tuned VLMs; its repositories remain private.

06 · Reactive simulation

Shear-driven mechanochemistry

My Ph.D. research used ReaxFF molecular dynamics in LAMMPS to resolve bond breaking, bond formation, and reaction pathways at sliding interfaces. I developed analysis tools for detecting reaction events, classifying products, and tracking where and how quickly chemistry occurred, alongside NEB calculations of pathway barriers.

The work reproduced the experimentally observed exponential dependence of reaction rate on shear stress and established a computational protocol for quantifying molecular deformation under load. Collaborations with experimental chemists connected the simulations to tribometry, Raman, SEM, and EDX measurements. The 2024 Scientific Reports paper was ranked in the journal's Materials Science Top 100.