Atomistic machine learning, open materials data, and a few tools around them.

I’m Ali, a PhD researcher at Entalpic and CentraleSupélec / CVN (Inria) in Paris. Each project below comes with something to play with.

Research

ICML 2026 · arXiv 2605.20581

TriForces

Augmenting Atomistic GNNs for Transferable Representations

Ali Ramlaoui, Alexandre Duval, Hannah Bull, Victor Schmidt,

Preserving the information a model needs when the chemistry changes.

streams for composition, structure and interactions: 3; streams for composition, structure and interactions

Paper ↗Code ↗

Two species, one structure. Colour identifies each atom’s type.The same atoms, seen through their nearest-neighbour bonds.One atom and its immediate neighbours. Select another to compare.

64 atoms, two species. Select one to inspect.

Three views of one structure.

Preprint · arXiv 2606.29975

Atompack

A Storage and Distribution Layer for Read-Heavy Atomistic ML Training Datasets

Ali Ramlaoui, Daniel T. Speckhard, Sagar Pal, Fragkiskos D. Malliaros,

A storage format designed around what the training loop actually reads.

faster shuffled reads than ASE LMDB: 96×
smaller files than ASE LMDB: −79%; smaller files than ASE LMDB

Paper ↗Code ↗Docs ↗

Training reads complete molecules in random order. The index points straight at each record in the memory-mapped file.Reading front to back. Contiguous formats all do well here; the gap narrows.Records are appended during construction, then the index is written and the spare header slot is flipped to commit.The same molecules on disk. Atompack stores each record as raw arrays with a compact index.

Atompack–
HDF5 SOA–
LMDB Packed–
LMDB Pickle–
ASE SQLite–
ASE LMDB–

molecules / s · 64 atoms · NVMe · 1 workermolecules / s · 64 atoms · NVMe · 1 workermolecules / s · built-in fieldsfile size relative to Atompack · lower is smaller

Atompack against five common formats, on one workload.

NeurIPS 2025 · AI4Mat Workshop · arXiv 2508.20875

LeMat-Traj

A Scalable and Unified Dataset of Materials Trajectories for Atomistic Modeling

Ali Ramlaoui, Martin Siron, Inel Djafar, Joseph Musielewicz,

Making fragmented materials calculations usable together.

atomic configurations: 120M+; atomic configurations
source databases, harmonized: 3; source databases, harmonized

Paper ↗Dataset ↗Data pipeline ↗

E − E_final (eV)

max force (eV/Å, log)

Loading a trajectory…

121.8M configurations. Hover a segment for the exact count.

PBE114.7M · 94.2%
PBESol6.4M · 5.3%
r²SCAN517k · 0.4%
SCAN181k · 0.1%

Materials ProjectAlexandriaOQMD

Relaxations from LeMat-Traj.

Software

Software · Apache-2.0

ssync

Taking an experiment from a local workspace to a cluster, and back.

ssync.ramlaoui.org ↗Code ↗

Type help, or watch the replay.

Launch, follow, read the logs.

More work

Software
Cavreno
Keeping remote development work within reach on iPhone and iPad.
cavreno.app ↗
Software
Commonthread
Reading coding-agent conversations without losing their original context.
Ask about it →
Data · 2026
LeMat-Rho
A high-fidelity r²SCAN charge-density dataset for materials modelling and electronic-structure analysis.
Dataset ↗Announcement ↗
Software · 2025 · Apache-2.0
LeMaterial-Fetcher
A concurrent pipeline for fetching, validating, harmonizing, and publishing materials data from heterogeneous sources.
Code ↗

Recent papers

All 8 papers →

About

I work on machine learning for materials. Most of my time goes into building architectures and training methods for atomistic models, and the datasets they learn from. I also build the tools I end up needing, like ssync for running jobs on clusters and Cavreno for keeping an eye on remote work from my phone.

email me