AI for Science × Materials Physics
Physics-Grounded ML & LLM Agents · Scattering Data · Disorder Modeling · Scientific Software
I build ML and LLM tools that turn complex scattering data into discovery — grounded in real materials physics, from sample preparation and neutron/synchrotron/STM measurements to physics-based modeling. My niche is owning the path from material and measurement design to reproducible Python/browser/LLM workflows that produce defensible, structure-property insight rather than black-box predictions.
Capability map
Sample → signal → model → software: a connected toolkit for turning hard materials data into decisions.
Synthesis & Sample Prep
Prepare crystals, polycrystalline samples, and thin films; tune synthesis routes; handle air-sensitive workflows; and screen sample quality before beamtime or surface measurements.
Proof: Flux/CVT growth, glovebox synthesis, MBE prep.
Research projects →Scattering & Metrology
Design neutron, synchrotron, and STM measurements, then convert diffraction, PDF, diffuse, inelastic, low-temperature, and high-pressure data into structure-property constraints.
Proof: ORNL/SNS; neutron, synchrotron, STM studies.
Publications →Modeling & Interpretation
Use DFT, phonons, Reverse Monte Carlo, Rietveld/magnetic refinement, and symmetry analysis to test mechanisms for disorder, magnetism, topology, and lattice dynamics.
Proof: Nature Communications, JACS, PRB/PRR studies.
Publications →Scientific Software & AI
Build Python/HPC and browser-first tools with Pyodide/WebGPU/React, plus LLM agents for analysis review, retrieval-grounded reasoning, local inference, and evaluation.
Proof: MATERIA, NEBULA3D, RMCProfile Workbench, rmc-phonon-dynamics.
Packages & tools →Scientific software I ship
Browser-first research tools for structure refinement and neutron/RMC/phonon analysis: local data, inspectable workflows, and reproducible outputs.
MATERIA
An AI-native workbench for crystal & magnetic structure refinement in the browser — single-crystal + powder, X-ray + neutron, nuclear + magnetic, reciprocal-space + PDF on one engine — which samples the posterior rather than only linearizing it, and exposes its pure core to LLM agents as 33 MCP tools.
NEBULA3D
Cleans 3D reciprocal-space neutron diffuse-scattering volumes and computes 3D-ΔPDF maps — the complete Python pipeline runs client-side via Pyodide, with an optional AI reasoning review where a local or cloud LLM grades the reduction quality.
RMCProfile Workbench
Dashboard for RMCProfile/STOG refinements — live monitoring, space-group detection, WebGPU KDE slices, plus a built-in LLM assistant (local via Ollama/LM Studio or cloud) that reasons over your run and watches convergence.
rmc-phonon-dynamics
Phonon band structures, DOS, animated 3D modes, and simulated INS spectra extracted directly from RMC ensembles — bands and the S(Q,E)-derived DOS now share one meV energy axis, so computed dispersion and measured spectrum read against each other. WebGPU compute shaders deliver the main ~100× speedup.
📌 Recent highlights
- [Software] MATERIA Workbench — Bayesian uncertainty and symmetry-mode refinement: the engine now samples the posterior — ensemble MCMC and gradient-based NUTS — instead of only linearizing it, matching the linearized esds to 1% on the Ni PDF golden. It also fits real-space distortion modes from an isotropy-subgroup tree, with a magnetic-PDF track validated against
diffpy.mpdf— all of it reachable by LLM agents through the 33-tool MCP surface. - [Publication] First-author Nature Communications (2026) study on Mn3Ga, revealing an intrinsic topological phase transition at room temperature driven by a magnetostructural transformation.
- [Software] scattering-ai-sdk (early development): an agentic AI layer for scattering science — modular agent skills, retrieval-grounded LLM reasoning, and evaluation harnesses, running fully offline with local models.
- [Software] Athanor (exploratory): a closed-loop prototype testing whether an LLM agent can help drive materials screening with physics-grounded surrogates, benchmarked against non-LLM baselines — an early direction I am actively exploring.
- [Software] Neutron diffuse scattering tools for 3D-ΔPDF analysis released: Developed a Python-based workflow for 3D-ΔPDF reconstruction and visualization, supporting analysis of local disorder and short-range correlations in complex materials.
- [Publication] First-author JACS (2024) study on kagome (Fe,Co)Sn, revealing coupling between short-range local disorder and a long-range antiferromagnetic transition.
- [Software] Released rmcph: a data-processing pipeline and GUI for calculating phonon spectra from total scattering measurements and RMC model ensembles, with integrated tools for phonon processing and visualization.