Athmane BenarousChemia Discovery · Machine Learning Intern · Feb – Aug 2026
ML pipelines and evaluation over 600K+ materials
30+ h → < 1 hpipeline runtime, after parallel S3 fetching and batched API calls
Work under NDA, described at the level of my CV: tools, scale and outcomes. No data, code or internal details.
Problem
Chemia predicts a range of materials properties with CGCNN (crystal graph convolutional neural network) models. Models have to be trained and evaluated, and their predictions backfilled across more than 600,000 materials. A full pipeline run took more than 30 hours.
Constraints
- Scale: 600K+ materials.
- Stack: crystal structure files in AWS S3, metadata and properties in MongoDB, compute on Azure ML.
What I built
- Azure ML pipelines for training, evaluation and automated backfill inference of CGCNN property models.
- Faster I/O: parallel S3 fetching and batched API calls.
- An evaluation layer: held-out benchmarks, a model card for every deployed model, and data-provenance scoring.
- A training dashboard (React, NestJS, MongoDB, PostgreSQL) with model comparison, space-group filtering and real-time per-step job monitoring.
Results
- Pipeline runtime from 30+ hours to under 1 hour.
- Automated backfill inference across 600K+ materials.
- A model card for every deployed model.
What went wrong and how I fixed it
A full run took more than 30 hours. Parallel S3 fetching and batched API calls brought it under 1 hour.
What I’d do next
- Track held-out benchmark scores per model version, so regressions show up before deployment.
- Report an uncertainty estimate with each prediction.