Tianhao Li

AI for Science

Tianhao Li

I build machine learning that makes scientific discovery data-efficient, reliable, and agentic.

My research adapts foundation models to out-of-distribution materials with minimal labels, evaluates them beyond standard error metrics, and builds LLM agents that ground hypotheses in literature and physical evidence.

Email me CV Google Scholar GitHub LinkedIn
Tianhao Li on a beach with a forested sea stack behind him
01

Research

Making scientific machine learning data-efficient and trustworthy, from adapting foundation models to evaluating them and building the data systems underneath.

Foundation models · Data selection 2024–now

Data-Efficient Adaptation of Foundation ML Interatomic Potentials

Universal potentials break down on ultra-complex disordered materials, where every label costs a DFT calculation. Reframing fine-tuning as data selection, a similarity-guided loop adapts CHGNet with about a hundred well-chosen structures, while random selection stalls and even degrades as data grows. The adapted model then drives synthesizability screening across millions of relaxations.

Manuscript in preparation
Benchmarking · Reliability 2025–2026

Evaluating ML Interatomic Potentials Beyond Structural Metrics

A benchmark of four universal MLIPs shows that getting structures right says little about getting electronic structure right, and deployment tests expose failures that static error metrics miss.

Paper
Surrogate models · Screening 2024–2025

ML-Guided Discovery of High-Entropy Fuel-Cell Catalysts

Physics-informed surrogates extend a high-throughput DFT campaign fifty-fold, and multi-objective ranking under uncertainty surfaces platinum-free candidates that balance activity, stability, cost, and supply risk.

Paper Code
Scientific data · Physics 2026

Physics-Based Reference Data for Automated Diffraction Analysis

A phonon-derived Debye–Waller temperature for computed XRD references cuts high-angle intensity errors by an order of magnitude, giving automated and ML phase identification a more faithful reference.

Manuscript in preparation
Data platforms · LLM agents 2023–now

Scientific Data & LLM Infrastructure

Built and run the high-throughput pipeline behind the lab's CHAOS database, its public search API, and LLM agent, and co-developed CHAOS-MatForge, an open dataset of millions of DFT-labeled structures for training interatomic potentials.

CHAOS Dataset
02

Experience

Jun–Sep 2026

Amazon

Applied Scientist Intern · Agentic AI for Scientific Discovery

Built an end-to-end AI co-scientist that carries a materials-discovery problem from the literature to the lab bench.

  • Architected a multi-agent system on Claude (Amazon Bedrock) that mines the literature, generates competing hypotheses, ranks them through tournament-style debate, and promotes the strongest to physics-based screening and bench validation.
  • Designed the evidence layer that makes agent output trustworthy: a literature agent that grounds every claim in verbatim source quotes, an independent auditor agent, and leakage audits.
  • Closed the loop between language models and physical evidence, connecting agent proposals to heat-transfer simulation and a calibrated bench campaign, as sole developer of the full application.

Multi-agent systemsRetrieval & groundingLLM evaluationAmazon Bedrock

2023–now

Johns Hopkins University

Graduate Research Assistant · Entropy for Energy Laboratory · Advisor: Prof. Corey Oses

Data-efficient ML for materials discovery: foundation-model adaptation, model evaluation, surrogate-driven screening, and the lab's scientific data and LLM infrastructure.

PyTorchGNN potentialsActive learningHPCDocker

03

Papers

Google Scholar

2026

  1. Structural Benchmarks Overlook Density-of-States Errors in Machine-Learned Interatomic Potentials APL Machine Learning
  2. Stabilizing Iodine in Pyrochlore: Toward New Nuclear Waste Forms Precision Chemistry

2025

  1. The Search for High-Entropy Fuel-Cell Catalysts Using Disorder Descriptors Nano Futures
  2. High Entropy Powering Green Energy: Hydrogen, Batteries, Electronics, and Catalysis npj Computational Materials
  3. Beyond the Four Core Effects: Revisiting Thermoelectrics with a High-Entropy Design Materials Horizons

2019

  1. Effects of H2–H3 Phase Transition Reversibility on Ni-Rich Layered Oxide Cathodes Frontiers in Chemistry

In progress

  1. From Training Data to Molecular-Dynamics Deployment: Transfer of Machine-Learning Interatomic Potentials in a Refractory High-Entropy Alloy Machine Learning: Engineering · under review
  2. Designing Ultra-Complex Structurally Disordered Materials via Machine Learning Interatomic Potentials In preparation
  3. A Debye–Waller Temperature for Computed Diffraction References In preparation
04

About

Education

  • Johns Hopkins UniversityPh.D., Materials Science & Engineering
  • Duke UniversityM.S., Materials Science & Engineering
  • Changsha University of Science & TechnologyB.S., Materials Science & Engineering

Honors

  • Amazon AI PhD FellowAmazon AI PhD Fellowship Program, Johns Hopkins

Service & teaching

  • Peer reviewerNature Communications; International Journal of Applied Glass Science (with advisor)
  • Teaching assistantGateway Computing: Python (EN.500.113)
  • Teaching assistantIntroduction to Computational Materials Modeling (EN.510.466/666)

Talks

  • Data-Driven Materials Modeling Workshop, JHU Talk · recording
  • ARPA-E Energy Innovation SummitPoster
  • ARPA-E Annual Fission MeetingPoster