A blue train crossing green fields below the snow-capped Djurdjura range — View of the Djurdjura from Tuvirett — southern side of Kabylie

Sami Chemlal

Applied Mathematics & Machine Learning · CentraleSupélec

Paris, France · formerly research intern at CMAP, École Polytechnique

samychemlal@yahoo.fr GitHub LinkedIn CV

About me

I am an engineering student in Applied Mathematics, Machine Learning & Statistics at CentraleSupélec (MSc-equivalent, 2023–2027), after preparatory classes (MPSI–PSI*) at Lycée du Parc in Lyon.

From March to August 2026 I was a research intern at CMAP, École Polytechnique, supervised by Rémi Flamary, Karim Lounici and Thibaut Germain. I worked on learning from populations of related systems: how to share structure across many dynamical systems or conditional distributions, so that a new one can be estimated from very little data. This led to two manuscripts, DOODL and MTL-CMO.

I also coach the Algerian national team for the International Olympiad in AI; my course notes are online.

Research interests

Scientific Machine Learning Optimization & Optimal Transport Operator Learning Learning for Physical & Dynamical Systems Multitask & Transfer Learning Representation Learning Generative Modeling Uncertainty Quantification

Award

Silver medal, AEFE International Chemistry Olympiad 2020

self-portrait, drawn by edge detection

News

Sep 2026Preprint on arXiv: Multi-Task Learning of Conditional Mean Operators (MTL-CMO).
Sep 2026Both manuscripts, DOODL and MTL-CMO, submitted to the ICLR 2027 main conference track.
Aug 2026End of my six-month research internship at CMAP, École Polytechnique.
May 2026Preprint on arXiv: Geometric Dictionary Learning of Dynamical Systems with Optimal Transport (DOODL).
Mar 2026Started a research internship at CMAP with R. Flamary, K. Lounici and T. Germain.
2025Started coaching the Algerian national AI-olympiad team (AOAI).

Publications

Preprints and submissions. Names in bold are me.

MTL-CMO

Multi-Task Learning of Conditional Mean Operators: Applications to Dynamical Systems and Uncertainty Quantification

S. Chemlal, T. Germain, R. Flamary, V. R. Kostic, K. Lounici

arXiv:2609.35429, 2026 · Submitted to ICLR 2027

A multi-task operator-learning framework for populations of conditional distributions. It learns function spaces φ and ψ shared across related tasks, and transfers in closed form to a new distribution from limited data, with statistical guarantees. Applications include metastable Langevin dynamics and turbulent plasma simulations.

Many dynamics. Shared spaces.

Concept animation · illustrative data

arXiv PDF

DOODL

Geometric Dictionary Learning of Dynamical Systems with Optimal Transport

T. Germain, S. Chemlal, R. Flamary, V. R. Kostic, K. Lounici

arXiv:2605.18276, 2026 · Submitted to ICLR 2027

Related dynamical systems lie near a low-dimensional manifold in spectral operator space. DOODL learns a dictionary of characteristic spectral dynamics with Riemannian optimization and optimal transport, and estimates a new system from short trajectories. In low-data regimes it gives one to two orders of magnitude lower operator-estimation error than estimating each system independently.

A dictionary of dynamics.

Concept animation · illustrative data · schematic geometry, 3 atoms for illustration

arXiv PDF

Application: dynamical systems

One example is turbulent plasma, which appears in both papers.

Real Tokam2D simulation data

Related research to understand my papers

Two papers by my co-authors that my work builds on directly.

Courses

I teach machine learning as a volunteer coach for AOAI, the Algerian national selection for the International Olympiad in Artificial Intelligence (IOAI). Here are some of the courses I wrote. Each one rebuilds a core idea from first principles and ends with a hands-on challenge. I add new ones regularly.

01

Courses on Probability and Optimization

The two views of learning: empirical risk minimization and maximum likelihood.

Markov, Chebyshev, law of large numbers
Cauchy–Schwarz, gradient descent and its convergence
GD vs SGD
MLE; challenge: Gaussian-mixture clustering

PDF

02

Linear Regression for ML

Covariance, linear regression and regularization, from geometry to real data.

Correlation as a cosine
Least squares, multivariate regression
Bias–variance tradeoff
Ridge & Lasso, Bayesian view; challenge: house prices

PDF

03

Transformers

From attention to modern language models, rebuilt from simple mathematical principles.

Embeddings, positional information
Self-attention and the 1/√d_k factor
Multi-head attention, FFN, residuals
Causal masking, next-token MLE, KV cache; challenge: NumPy attention layer

PDF

Curriculum Vitae

Download CV (PDF) sami-rayan.chemlal@student-cs.fr

Education

2023 – 2027
CentraleSupélec: Diplôme d'Ingénieur (MSc-equivalent)

Applied Mathematics, Machine Learning & Statistics (MDS-SDI) · GPA 3.92 / 4.00
- Modern deep learning, transformers, NLP, diffusion & generative models, representation and transfer learning, reinforcement learning, statistical learning, optimization, optimal transport.
2020 – 2023

Lycée du Parc: CPGE MPSI–PSI*

GPA 4.00 / 4.00 · Ranked 4th in the CentraleSupélec international entrance examination

Research experience

Mar – Aug 2026
Research Intern: Scientific Machine Learning

CMAP, École Polytechnique, Palaiseau · Supervisors: R. Flamary, K. Lounici, T. Germain
- Developed MTL-CMO, a multi-task operator-learning framework for populations of conditional distributions, with closed-form transfer to new distributions from limited data and statistical guarantees.
- Co-developed DOODL, a geometric dictionary-learning framework for populations of dynamical systems combining spectral operator representations, Riemannian optimization and optimal transport.
- Applied both to uncertainty quantification and stochastic dynamics (metastable Langevin dynamics, turbulent plasma simulations): conditional statistics, system identification, parameter recovery and regime characterization.
- One to two orders of magnitude lower operator-estimation error in low-data regimes; two manuscripts submitted to ICLR 2027.

Research projects

2026
Predictive Corruption Geometry for Generative Modeling of Physical Fields

Independent research project · Generative modeling, spectral methods, PyTorch, GPU/SLURM
- Studied corruption geometry as an inductive bias for generative models of physical fields, via a mode-wise Fourier analysis of the denoising risk.
- Derived analytical criteria predicting when PDE-induced and covariance-aligned corruptions are complementary, and built a dual-geometry representation from them.
- Evaluated on Kolmogorov turbulence, decaying 2D turbulence and Cahn–Hilliard dynamics. The analytical selector matched the choice made by exhaustive validation and improved on a tuned spectral baseline.
2025
Robust Pattern Classification for Environmental Time Series

CentraleSupélec / L2S Laboratory · Deep learning, time series, robustness
- Built synthetic benchmarks with controlled morphology, noise, missing segments and scale shifts.
- Compared ANN, CNN, LSTM and Transformer models under severe noise, partial observations and changes in pattern geometry.
- Developed an online sliding-window pipeline to detect and classify motifs in continuous signals.
2026
QRT Data Challenge: Cross-Sectional Return Prediction

Machine learning for financial time series
- Momentum, volatility, liquidity and relative-value features; ensembles of tree-based models. Ranked 36th out of 1,447 (top 2.5%).

Industry experience

Aug 2025 – Feb 2026
Quantitative Research Intern: Front Office

Crédit Agricole CIB, Paris
- Statistical models for credit-event and default prediction (linear, logistic, regularized regression; calibration, out-of-sample validation).
- Monte Carlo models of correlated defaults, with dependence through covariance structures and Cholesky factorization.
- Repayment and prepayment distributions; KS, χ², t- and F-tests, heteroscedasticity, multicollinearity, weighted least squares.

Teaching

2025 – present
AI Olympiad Coach (volunteer)

AOAI, Algerian national selection · Remote
- Probability, statistical learning, optimization and deep learning. See the courses →
2023 – present
Mathematics Tutor

CPGE & undergraduate mathematics · Paris
- Real analysis, topology, measure and integration, differential calculus, linear algebra, reduction theory, Euclidean spaces. Students went on to CentraleSupélec, Télécom Paris and HEC Paris.

Skills

Mathematics: Probability, stochastic calculus, harmonic analysis, optimal transport, convex & non-convex optimization, dynamical systems, operator learning
Machine learning: Scientific ML, transformers, diffusion & flow models, representation learning, multitask & transfer learning, reinforcement learning, NLP
Programming: Python, C#, R, MATLAB, SQL · PyTorch, JAX, TensorFlow, scikit-learn, pandas, NumPy, SciPy, POT · CUDA, SLURM, Git
Languages: French (C2), Kabyle (C2), English (C1)

Sami Chemlal

About me

News

Publications

Multi-Task Learning of Conditional Mean Operators: Applications to Dynamical Systems and Uncertainty Quantification

Geometric Dictionary Learning of Dynamical Systems with Optimal Transport

Application: dynamical systems

Related research to understand my papers

Neural Conditional Probability for Uncertainty Quantification

A Spectral-Grassmann Wasserstein Metric for Operator Representations of Dynamical Systems

Courses

Courses on Probability and Optimization

Linear Regression for ML

Transformers

Curriculum Vitae

Education

CentraleSupélec: Diplôme d'Ingénieur (MSc-equivalent)

Lycée du Parc: CPGE MPSI–PSI*

Research experience

Research Intern: Scientific Machine Learning

Research projects

Predictive Corruption Geometry for Generative Modeling of Physical Fields

Robust Pattern Classification for Environmental Time Series

QRT Data Challenge: Cross-Sectional Return Prediction

Industry experience

Quantitative Research Intern: Front Office

Teaching

AI Olympiad Coach (volunteer)

Mathematics Tutor

Skills