Ziqi Wang

PhD candidate, Mechanical Engineering and Scientific Computing, University of Michigan

I work at the intersection of computational materials science and agentic AI systems. I build multi-agent systems that run the physics simulations behind materials discovery autonomously on supercomputers. Because a correct answer can hide fabricated reasoning, I also build the verification systems that catch agents fabricating results. I was selected for Anthropic’s AI for Science program. I am seeking applied research scientist roles and graduate in summer 2026.

Headline results

  • 0.2% error on adsorption-energy differences between preferred sites, using the same exchange-correlation functional (CO/Pt(111))
  • Below 1% error on lattice-constant calculations across 27 elemental bulk materials spanning 3 crystal structures (Sol27LC)
  • Up to 3x token reduction for the lightweight DREAMS framework compared to baseline frameworks
  • Worker error rate and judge success rate are currently under benchmarking
  • DREAMS-OER catalyst exploration is currently running; result analysis is preliminary

Projects

project thumbnail

DREAMS

a hierarchical multi-agent AI framework that runs density functional theory simulations end to end, validated at human expert accuracy on standard benchmarks
project thumbnail

Transparency, Provenance, and Trustworthiness

provenance records, fabrication-proof verification, and cost accounting that make agent reasoning inspectable
project thumbnail

DREAMS-OER

an autonomous agent that screens a materials database for oxygen-evolution catalysts over a multi-hour campaign, tracking hundreds of interdependent simulations without losing the thread

Resume

Trustworthy scientific LLM agent researcher. First author of DREAMS.

Open full resume (PDF) ↗

Education

2021 – 2026
Ph.D., Mechanical Engineering and Scientific Computing
Carnegie Mellon University and University of Michigan
Advisor: Prof. Venkatasubramanian Viswanathan
2016 – 2021
B.S., Mechanical Engineering and B.S., Computer Science
University of Michigan and Florida Institute of Technology

Research experience

arXiv 2025
DREAMS first author
Fully autonomous LLM agent that runs expert-level materials simulations
  • Reached human-expert accuracy on multi-step autonomous workflows: reproduced a long-debated surface-chemistry benchmark (CO/Pt(111)) to within 0.2% and matched expert structures on all 27 systems of a crystal benchmark, where a single-agent baseline failed every run and a multi-agent baseline missed by 389%.
  • Canvas: a shared communication-and-memory scheme with report-based history compression for long-horizon runs; benchmarked against single-agent and multi-agent baselines on accuracy, success rate, and token cost, using up to roughly 3x fewer tokens on long-horizon tasks.
  • Deterministic and LLM-based safety guards validate every parameter against customizable rule levels and user-defined sensitive parameters; tuned on an adversarial harness of 145 scenarios from real DFT runs, best judge at 1 false positive and 3 false negatives (n=145) across five frontier LLMs.
  • Full traceability: every result registered with its inputs and outputs, verified by reference ID, and linked to its sources through a provenance DAG, each claim backed by an ID-annotated reasoning chain, enabling recursive debugging of long runs.
In preparation
DREAMS-OER
Autonomous agent for open-ended catalyst discovery over a roughly 380,000-material space
  • Scaled the agent from a single fixed task to open-ended search over roughly 380,000 candidate materials (Google DeepMind's GNoME set, DFT via VASP), choosing what to study next (material, surface, site, three intermediates) from accumulated evidence rather than a fixed plan. Demonstrated on behavioral runs; production screening in progress.
  • Grounded decisions in retrieval-augmented (RAG) literature results and a multi-level relational experiment log that enables efficient information retrieval and progress tracking across hundreds of partial, interdependent studies.
  • Gave the agent live time- and compute-budget awareness to submit work opportunistically, keep the cluster busy while reasoning, and re-plan under pressure, completing a bounded 7-hour study without overrunning.
  • Caught workers specification-gaming a queue-occupancy target, submitting jobs to satisfy the metric without analyzing results; fixed it structurally with a disposition gate making engagement with every completed result a verifiable predicate before rest, plus a runaway-loop guard.
In preparation
Alloy Thermodynamics and Phase Behavior with Machine-Learning Interatomic Potentials
  • Built a committee-based active-learning loop to fine-tune universal MLIPs (MACE, NequIP, GRACE, MatterSim) into low-cost surrogates for the Li-Mg alloy system, cutting prediction error by roughly 5x versus off-the-shelf models; stress-weighted training gave the largest gains on hard second-derivative properties.
  • Ran LAMMPS molecular dynamics melting and solidification campaigns with MLIPs and classical potentials to locate melting temperatures across composition for binary and ternary lithium alloys.
  • Mapped solid-phase boundaries of binary alloys with canonical and semi-grand canonical Monte Carlo, using cluster-expansion models and LAMMPS with MLIPs and classical potentials.
  • Computed free energies by thermodynamic integration to an Einstein crystal; constructed phase diagrams via thermodynamic integration, metadynamics, heat capacity, and moving-interface methods.
  • Quantified non-linear excess effects in stiffness and atomic-transport barriers across alloy composition.

Publications and preprints

2025
Z. Wang, H. Huang, H. Zhao, C. Xu, S. Zhu, J. Janssen, V. Viswanathan. "DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation." arXiv:2507.14267 (2025).
2025
A. M. Yao, S.-J. Kim, V. D. Veeraraghavan, Z. Wang, S. Gasilov, A. Kastengren, Y.-M. Chiang, P. Fenter, V. Viswanathan. "Effective Li-Ion Transport Quantification in Composite Cathodes for All-Solid-State Batteries via Multiscale Modeling and Experiments." Chem. Mater. 37, 9260-9267 (2025).

Technical skills

Agentic AI / LLM
Multi-agent system design, harness design, SKILL design, guardrail and provenance systems, agent evaluation and benchmarking, RAG, LangGraph, LangChain
Machine learning
PyTorch, machine-learning interatomic potentials (MACE, NequIP, MatterSim, GRACE, SchNet), graph neural networks, active learning
Atomistic simulation
DFT (Quantum ESPRESSO, VASP), ASE, molecular dynamics (LAMMPS), Monte Carlo, cluster expansion, thermodynamic integration, nudged elastic band
Computing
Python, C++, HPC / SLURM, Git

Honors and awards

2026
Selected for the Anthropic AI for Science Program
2019 – 2021
University Honors, University of Michigan