Research focus
Physics-informed neural networks, graph learning, and quantum-informed modeling for property prediction.
PINN
GNN
Quantum-informed
Explore
Materials Science Research
Updates and demos from my materials science research at MBZUAI on physics-informed and data-driven models for molecules and crystals.
Upcoming demos
Interactive examples, baselines, and datasets as experiments mature. Collaborators welcome.
Live simulation
Atomic Interaction Lab
A molecular-dynamics engine running in the browser: Lennard-Jones atoms, a Langevin thermostat, and the force field they generate drawn live on top of them. Ten starting states from a cold crystal to a dilute vapour, with g(r), energy, the Maxwell–Boltzmann speed distribution and the diffusion coefficient measured from the same trajectory. Drag an atom, melt a lattice with the heat brush, and export the trajectory as CSV.
Lennard-Jones
BAOAB Langevin
Force fields
Phase transitions
Interactive demo
The (ML) force field family tree
A time map of how atoms became numbers — every representation trunk from Lifson’s 1968 consistent force field to the universal foundation models of 2026, the benchmarks that rank them, the groups that build them, and the problems still open.
Models — how atoms became numbers
Lanes are representations, not architectures: what a method feeds the regressor determines what it can and can’t learn, and that choice is what actually forks the tree. The horizontal axis is compressed before 2015 and stretched after — that skew is the story. Foundation models carry a red ring.
atom-centred
message passing
attention / transformer
universal / foundation model
hover → lineage · click → detail
↔ scroll horizontally
All models
The same set, as reference cards. Filters above apply here too.
Benchmarks — and who currently holds them
A model is only as meaningful as the set it was measured on. Lanes here are domains; each dataset carries its level of theory, its size, and the models that top it. Numbers are as published — read them as pointers to leaderboards, not as a settled ranking.
dataset
benchmark with leaderboard
click → SOTA table
↔ scroll horizontally
All benchmarks & datasets
Groups, labs, and industry teams
Academic groups and industry labs actively shipping methods, with the models they’re known for — a target list for collaborations, internships, and conference corridors.
State of the field, mid-2026
Open problems, ranked by how crowded they are
Fourteen directions with real headroom, each with the papers that define it and the groups working on it. Underserved and nascent are where a thesis can still claim territory; crowded means racing well-funded teams on their turf.
The stack you’ll actually type into
Venues, journals, schools, and reviews
Where this work gets published and argued about, plus the review papers worth reading first.
Compiled 2026-07-17 by an agent research fleet; all 710 outbound links were resolved over HTTP and returned live. Entries that could not be fully confirmed are marked unverified. The field moves weekly and leaderboards turn over — treat every number as a pointer to a leaderboard, not a citation. Lineage taxonomy and trunk colours after the “(ML) force field representation family tree” figure.