Xiao Huang
Portrait of Xiao Huang wearing a navy suit.
01 / RESEARCH & ENGINEERINGXiao Huang
Xiao Huang seated at a grand piano in a recording room.
02 / LIFE BEYOND THE LABAt the piano

AI FOR SCIENCE / RESEARCHER & ENGINEER

Hi, I’m
Xiao Huang.

I’m from Beijing, China. I studied Chemistry at Emory University, with a minor in Music. Today, I’m a Chemistry PhD candidate at Duke University, concurrently pursuing a master’s in Electrical & Computer Engineering.

I connect molecular science, machine learning, and software engineering—from modeling electron transfer in proteins to building scientific AI agents and designing new molecules. In 2026, I was a Software Engineering PhD Intern at Google Cloud.

Meet my research Download CV (public PDF)
ROLES OF INTEREST

Research Engineer / Applied Scientist / Software Engineer

AI for Science · Scientific AI Agents · ML + Chemistry

01 INDUSTRY EXPERIENCE

What I built
in industry.

Cloud incident diagnostics at Google and molecular simulation software at Deep Potential.

Google Cloud

AI Infrastructure team

Software Engineering PhD Intern · May-Aug 2026 · Kirkland, WA

AI INFRASTRUCTURE + ML

Unsupervised ML Anomaly Detection & User Behavioral Fingerprint Modeling for Google Cloud Infrastructure Traffic

  • Architected a data aggregation pipeline for the GCE metadata storage infra over Spanner, achieving a 50,000× size reduction to fuel a SQL backend unsupervised Machine Learning (ML) Anomaly Detection Model.
  • Engineered AI infrastructure for Time-series Anomaly Detection by fine-tuning a Bayesian Time-Series algorithm, establishing robust models and metrics to monitor and describe clients’ behavioral metrics.
  • Developed a full-stack architecture featuring a Python API layer and TypeScript frontend to support various resolution queries and deliver a comprehensive interactive dashboard view.
  • Pioneered a system that is AI Agent native supportive by developing a custom SKILL.md.
  • Productionized and deployed all project components, now running reliably in production.
  • Drove major business impact by reducing debugging time for on-call engineers solving outages and bug tickets, attracting engineers to learn and apply the pipeline to broader business logic.

The aggregation preserves the information needed for behavioral analysis and anomaly investigation, rather than every original raw-log record.

Technical implementation: aggregation, unsupervised ML, and full-stack architecture
  1. 01 / RAW LOGSTransaction metricsConsolidate actions and attempts
  2. 02 / PRE-AGGREGATIONBehavioral fingerprintsClient, caller, region & time window
  3. 03 / SQL ANALYSISBayesian time seriesTable-valued function on aggregates
  4. 04 / INVESTIGATIONInteractive drill-downPython JSON API + TypeScript / Canvas
Pre-aggregate logs and control cardinality

Consolidated action and attempt records into transaction metrics, then grouped them by client, caller, region, and fixed time windows. Stored these derived fingerprints in queryable aggregate tables, limiting high-cardinality grouping dimensions while retaining representative transactions for investigation. Split historical backfills into bounded time-range batches, ran batches in parallel, and used off-peak execution to manage shuffle-memory and I/O pressure.

Engineer interpretable behavioral signals

Worked with on-call engineers to define features such as waiting-time fraction: total idle time divided by total transaction duration within each window. Integrated and tuned an unsupervised Bayesian structural time-series model with seasonal baselines and pre- and post-event context. Evaluated abrupt spikes and gradual drift relative to each client’s history, validating sensitivity and false positives against human-reviewed examples.

Run model inference inside SQL

Integrated the model as a SQL table-valued function (TVF), keeping analysis alongside the aggregate data instead of exporting full datasets to Python. Precomputed aggregates support repeated investigations without rescanning all raw logs. Balanced aggregation granularity, query cost, and reliable batch completion against interactive latency and data freshness.

Separate the API from the dashboard

Built a Python JSON API that maps time-window, resolution, and filter selections to parameterized SQL query templates, then returns compact results to a modular TypeScript dashboard. Custom Canvas charts support grouping and drill-down from behavioral shifts to representative transaction context. The browser communicates through the API rather than querying the database directly.

Data freshness and engineering trade-offs

The deployed system used scheduled batch aggregation for retrospective diagnosis. The follow-on design targets a maximum data lag of approximately seven hours, making data from about seven hours ago and earlier available for investigation. This is a planned refresh capability, not real-time alerting. The tool supports human investigation rather than automatic incident remediation.

To limit model-evaluation cost, I designed a candidate-screening stage with Z-scores for abrupt spikes and Mann–Kendall tests for gradual trends, followed by Bayesian analysis. This is a proposed extension to the deployed SQL-native detector.

  • Unsupervised ML
  • Bayesian time series
  • SQL TVFs
  • Python API + TypeScript
  • Parallel aggregation
  • AI Agent SKILL.md

Deep Potential Technology

AI for Science Research Intern · Feb-Aug 2022 · Beijing, China

MOLECULAR SIMULATION

Free-energy calculations for covalent drug design

  • Developed Covalent Uni-FEP, a high-accuracy free energy perturbation (FEP) algorithm for covalent drug performance evaluation, benchmarking binding affinity errors within 1 kcal/mol (chemical accuracy).
  • Automated complex inhibitor–protein topology setup and molecular dynamics (MD) preparation in GROMACS utilizing Python and Bash.
  • Accelerated large-scale free energy computations by engineering a domain decomposition strategy with multi-GPU support and an RDKit-based reaction tracking pipeline.
  • Authored core patent documentation and contributed production-quality code to the commercial software release.
Technical implementation: simulation setup, free-energy analysis, and debugging
Automate difficult molecular setup

Used Python and Bash to prepare inhibitor-protein topologies and GROMACS simulations, covering ligand alignment, force-field generation, perturbation setup, and molecular-dynamics preparation.

Make free-energy workflows practical

Worked on domain-decomposed, multi-GPU calculations and RDKit-based reaction processing. Investigated energy and trajectory inconsistencies across compute environments and supported batch simulation workflows.

Diagnose scientific failure modes

Examined convergence and sampling problems through MBAR energy decomposition, trajectory analysis, and binding-pocket hydration analysis. Fixed analysis scripts and contributed code through review.

Translate research into software

Authored algorithm and workflow sections of patent documentation and contributed production-quality code for the commercial Hermite Uni-FEP platform. Documented preparation and debugging procedures for handover.

The approximately 1 kcal/mol benchmark applies to selected test cases; accuracy depended on the system and sampling quality.

  • Python + Bash
  • GROMACS
  • RDKit
  • FEP + MBAR
  • Multi-GPU / Slurm
  • MDAnalysis

02 RESEARCH COORDINATES

Small molecules.
Big computational questions.

I use physics, machine learning, and AI to understand how molecules work—and engineer tools that help us design new ones.

01 / MOLECULAR PHYSICS
Electron transfer through a protein complexA conceptual protein surrounds linked molecular sites. An electron moves from site to site, with two possible routes. Geometry and energy determine the transfer rates.electron

How do proteins move electrons?

Think of a protein complex as a molecular relay: electrons hop between embedded sites. I model how distance, energy, and competing routes control where they go—and how light can drive charge separation.

The challenge: many interacting sites, different timescales, and a rapidly growing space of possible states. A “hole” is a missing electron; its movement is another way to describe charge transfer.

  • Electron-transfer theory
  • Protein design
  • Kinetic models
02 / MACHINE LEARNING + DE NOVO DESIGN
Language models and reinforcement learning for molecular designA chemical language model proposes a molecular structure. A computational reward scores it and guides the next round of generation. The molecular graph is illustrative, not a specific compound.ChemicalLLMScore &rewardRL feedback

Design new molecules, then learn.

De novo means designing from the start. I combine chemical language models with reinforcement learning to propose small molecules and improve their predicted properties through repeated feedback.

For EGFR drug design, the reward balances predicted potency, drug-likeness, synthetic accessibility, and novelty. These are computational candidates—not experimentally validated treatments.

  • LLMs + reinforcement learning
  • De novo design
  • XGBoost + SHAP
03 / SCIENTIFIC AI AGENTS
An AI agent connected to scientific computationA scientific question leads to reasoning and a proposed model edit, scientific validation, then a solver calculation. Numerical results return as evidence for the next step.Reason &proposeValidatethe modelRun thephysics solverResults → evidence → next question

AI that reasons—and calculates.

A scientific assistant should do more than explain. In Agentic MEKinetics, it can propose model changes, pass them through validation, and call a real physics engine to test a scientific question.

The challenge: connecting natural-language intent to explicit model parameters and reproducible numerical evidence. The demonstrated workflow is being evaluated more broadly.

  • Tool-using AI
  • Scientific validation
  • Reproducible results
04 / PARALLEL SCIENTIFIC COMPUTING
Parallel computation for exploring many scientific modelsOne parameter study distributes independent simulations across three workers. Their results are combined for analysis. This is a conceptual workflow, not a speed benchmark.ParameterstudyWorker 1Worker 2Worker 3Combine & analyze

Explore more possibilities, together.

A single simulation is only the beginning. I distribute independent model runs across workers, use sparse numerical methods, and organize parameter searches to explore many possible molecular designs.

The challenge: controlling memory costs and coordinating batches while keeping every result tied to its model and settings. Physics provides the model; computation makes exploration practical.

  • Parallel algorithms
  • Sparse solvers
  • Slurm / HPC

Conceptual diagrams for orientation—not specific molecular structures, measured data, or performance benchmarks.

03 SELECTED RESEARCH & PROJECTS

Models that explain.
Software that tests.

My research spans electron transfer in proteins, generative molecular design, and machine-learning potentials for materials. Each project connects a physical question to a testable computational model.

01 / PROTEIN DESIGN + INTERPRETABLE MLFirst author · 2025 & 2026

Designing proteins that
route light-generated charge.

Proteins move electrons through chains of molecular sites. I study how light-generated electron vacancies, or “holes,” can reach two destinations with different oxidizing strengths. The design challenge is to favor useful transfer routes while suppressing short circuits that waste the stored energy.

In our ACS Central Science study, we combined molecular dynamics, quantum-chemical calculations, kinetic master equations, and Bayesian optimization to study light-driven hole bifurcation. An illustrative model approached 100% simulated bifurcation yield.

The follow-up PCCP study examined 130,000 theoretical network configurations. An XGBoost classifier achieved 96.2% accuracy on 26,000 held-out simulated networks. SHAP analysis identified how site distances and electrostatic repulsion affect performance, including a trade-off between successful charge separation and energy efficiency.

What the model establishes

The network dataset varies physical distances and redox potentials; it is not a library of 130,000 protein sequences or atomistic structures. Atomistic calculations in the ACS study estimated about 1.6 eV of repulsion between the initially generated holes. The papers provide computational design rules, not experimental proof of a working light-powered protein.

  • Master-equation kinetics
  • DFT / MD / QM/MM
  • XGBoost + SHAP
  • Bayesian optimization
ACS paper · 2025 (opens in a new tab)PCCP paper · 2026 (opens in a new tab)

THE SCIENCE, IN THREE STEPS

A light-powered
molecular circuit.

A protein holds tiny molecular relay sites in place. The question: can we guide two charges to two different destinations?

01
Light creates two “holes.”

Two light-driven steps create two holes: missing electrons on neighboring sites.

Two holes travel through a designed protein along separate routes Two light-driven steps create two holes on neighboring molecular sites, shown as plus signs. Hole repulsion makes the first departing hole more strongly oxidizing than the second. Route A leads to the stronger-oxidizing endpoint; route B leads to the weaker-oxidizing endpoint. A crossed-out dashed connection represents an unwanted shortcut. The shaded outline represents the protein scaffold, not an actual protein structure. Two nearby sitesProtein scaffold = molecular support ++ Avoid shortcuts AB

Hole = missing electron · Arrows show hole transfer

02
One hole, then the other.

The nearby holes repel each other. The first leaves with a stronger ability to remove electrons from other molecules—its oxidizing power. The second leaves with less.

AStronger oxidizing power
BWeaker oxidizing power
03
Design the useful routes.

I use physics simulations and machine learning to find site spacings and energies that guide one hole to each destination, avoiding wasteful shortcuts.

Proposed protein design · simplified computational model, not an experimentally demonstrated device. Electrons move in the opposite direction to holes.
02 / GENERATIVE AI + ANTICANCER DRUG DISCOVERYCorresponding author · 2026

Teaching a language model to design anticancer drug candidates.

I co-developed a ChemBERTa and reinforcement-learning pipeline to design small-molecule anticancer drug candidates targeting EGFR. Starting with 11,394 curated ChEMBL molecules, we adapted the chemical language model through masked-language training, then used a REINFORCE-style policy gradient to guide generation.

A multi-objective reward combines Random Forest-predicted potency, drug-likeness, synthetic accessibility, and novelty. RL increased mean composite reward from approximately 0.34 to 0.50, about 45% above the adapted baseline. A 1,000-molecule sample was fully valid under RDKit and contained no duplicates.

Evaluation and limits

Docking and nearest-neighbor comparisons assessed predicted binding and structural differences from known molecules. The improved reward came with a trade-off in predicted ease of synthesis. The potency model used a random train/test split, and the candidates have not been validated in biochemical or cellular experiments. The reported reward increase is not a measured increase in drug efficacy.

  • ChemBERTa / Hugging Face
  • PyTorch + REINFORCE
  • RDKit + QSAR
  • AutoDock Vina
Read the paper (opens in a new tab)Code (opens in a new tab)
03 / ATOMISTIC MACHINE LEARNINGAccepted · awaiting publication

Testing AI force fields
for battery materials.

How reliably can an AI model predict atomic forces in a new environment? I co-developed a DFT-to-ML workflow for lithium on pristine, defective, and silicon-decorated graphene, building a 273-configuration DFT dataset and fine-tuning an equivariant MACE model.

On an independently defined test of 25 DFT configurations, fine-tuning reduced pooled force RMSE by 40.6%, from 450.0 to 267.5 meV/Å. Predictions improved across all five structural families and in 23 of the 25 configurations.

Transferability and my contribution

Structure-resolved tests exposed remaining force errors near highly strained carbon vacancies, showing where additional quantum-mechanical training data are needed. These results test local model transferability; they do not establish battery capacity, cycling performance, or diffusion coefficients. I am an equal-contribution and corresponding author, with shared methodology, software, analysis and writing, plus supervision.

Accepted by Computational Materials Science; awaiting formal publication. No public article link is available yet.

  • MACE / equivariant GNNs
  • Transfer learning
  • DFT / VASP
  • ASE + LAMMPS
Publication details
04 / SCIENTIFIC AI AGENTSWorking research software

Agentic MEKinetics

Reason about a model. Run the calculation.

I built a kinetics workbench that calculates how electrons and holes move through molecular networks. A Python simulation engine serves the command line, an interactive TypeScript dashboard, and a tool-using scientific AI assistant.

  • Natural-language model edits pass through scientific validation before a solver runs.
  • Shared calculation tools preserve model settings and reproducible result bundles across interfaces.
  • Grid scans, Bayesian parameter searches, sparse solvers, and independent parallel runs make larger studies practical.
AskValidateSimulate

A model-edit and simulation workflow has been demonstrated. Broader agent evaluation is ongoing.

  • Python + TypeScript
  • Numerical methods
  • Scientific agent tools
Discuss the project Private repository
05 / SCIENTIFIC SOFTWARECo-author · 2022

AutoSolvate:
from solvent to simulation.

I co-authored AutoSolvate, an open-source Python toolkit for preparing solution-phase quantum-chemistry calculations. It automates solvent-box construction, force-field and charge preparation, molecular-dynamics sampling, and extraction of molecular clusters through Python, command-line, and graphical interfaces.

The paper demonstrated the workflow on 166 redox couples, obtaining converged reorganization-energy results for 151 systems. A model trained on simulation-generated data predicted solvent-solute spacing in acetonitrile (MeCN) with a 0.09 Å mean absolute error on 83 held-out systems.

This project connects reproducible simulation setup with data generation for machine learning in chemistry.

  • Python
  • PACKMOL + AmberTools
  • Quantum chemistry
  • SOAP + gradient boosting
Read the paper (opens in a new tab)Code (opens in a new tab)

06 / CONNECTING THEORY TO BIOLOGY

From individual electron hops
to biological function.

I co-authored a review connecting electron-transfer theory to energy conversion, bacterial nanowires, oxidative-stress protection, and charge movement through DNA. My contributions include original-draft writing, visualization, and review and editing.

Read the review (opens in a new tab)

04 ABOUT

A scientist’s questions.
An engineer’s mindset.

I’m a PhD candidate in Chemistry at Duke University, also pursuing an M.S. in Electrical and Computer Engineering. My work connects physical models, machine learning, and the software that makes them useful.

I move between molecular systems and engineering systems: electron transfer in proteins, generative molecular design, scientific AI agents, and cloud observability. I enjoy working through both the scientific question and the implementation.

Outside research, I’ve played piano for over 20 years. Music was part of my education at Emory and remains a part of my life.

More on LinkedIn (opens in a new tab)Duke profile (opens in a new tab)

BEYOND RESEARCH

Another kind
of keyboard.

20+ years at the piano. A music minor at Emory. And a practice that remains part of my life alongside science.

Liszt · Gnomenreigen

Spring 2026 recital
Watch on YouTube (opens in a new tab)

05 SKILLS & TECH STACK

The working toolkit.

From physical models and learning algorithms to the systems that run them.

/ 01

Machine learning

  • PyTorch · Hugging Face
  • LLMs · Reinforcement learning
  • XGBoost · SHAP · QSAR
  • Bayesian optimization
  • Time-series anomaly detection
/ 02

Scientific computing

  • Charge-transfer kinetics
  • DFT · MD · QM/MM
  • Free-energy perturbation
  • RDKit · AutoDock Vina
  • GROMACS
/ 03

Software engineering

  • Python · Java · SQL
  • TypeScript · Bash · R
  • Distributed systems · GCP
  • Data pipelines · API development
  • Git · Docker · CI/CD
/ 04

AI agents for coding & research

AI coding agents in my development workflow, alongside tool-using agents for scientific computation.

  • Claude Code
  • Codex
  • OpenCode

AI coding agents · Tool use · Custom SKILL.md workflows

06 SELECTED PUBLICATIONS

Ideas, in the literature.

Google Scholar (opens in a new tab)

Accepted

Computational Materials Science · Awaiting formal publication

Machine-learning interatomic potentials for battery materials: Defect-dependent lithium adsorption and local transferability on graphene

Yuhan Sun, Xiao Huang (Corresponding Author) · Equal contribution

Accepted manuscript; DOI, volume and pages will be added when available.

07 LET’S CONNECT

What can we
discover next?

I’m interested in Research Engineer, Applied Scientist, and Software Engineer opportunities—especially in AI for Science, scientific agents, and ML for chemistry.

START A CONVERSATION

xhsean@outlook.com xiao.huang@alumni.emory.edu

Download my CV (public PDF)