Stories

Every story follows the same plan: the finding, the question, the theory in plain words, where it sits on the complexity ladder, an example in figures, what it brings and its limits. Technical details and code are on the methods site.

Compute directly · Hybridize · Learn frugally · AI in its place · Demonstrate trust

Compute directly

When the link is known, compute it with an explicit model, calibrated and validated on data; learning is added only for a demonstrated benefit.

BED per fraction for the rectum, LQ and LQL

Compute directly · Medicine

Compute, don't learn: radiobiological dose equivalence

One fraction of 15 Gy to the rectum: 72.7 Gy under LQ, 59.4 Gy under LQL.

One weekday: fixed base, auxiliaries, lighting and ventilation stacked

Compute directly · Energy

Simulate what you cannot yet measure: road tunnels

One year at 10-minute resolution in about 150 ms, with its limits stated.

Illustration du cours : indicateur F en fonction de l’horizon de 1 à 12 heures, sCV fixe et mémoire restante. Figure pédagogique distincte de la validation sur 68 stations.

Compute directly · Energy

How forecastable is the sun?

sCV and F describe a site before any model is fitted.

Hybridize

When the model falls short, complete it: keep the model, learn only what is missing.

Thyroglobulin over 80 months for 1, 2, 4 and 8 sessions, refractory and responding profiles

Hybridize · Medicine

Thyroid cancer and radioiodine: model first

Marie Fusella Giuntini and Dominique Barbolosi: three equations, 50 patients, a hybrid twin in progress.

Étude BLEND, figure 6 de la version auteur : erreurs nRMSE et nMAE moyennes sur 68 stations SIAR, horizons de trente à trois cent soixante minutes, variantes BLEND et modèles de référence.

Hybridize · Energy

Better persistence, without training a neural network

Symmetry-constrained forecasting: the structure of the signal written into the reference.

Learn frugally

When the useful relations are not known well enough, evaluate what the data can teach, with only as much complexity as pays.

Étude MIMO-ELM, figure 1 de la version auteur : mesures et prévisions à une heure pour sept signaux énergétiques corses, environ cent heures en hiver.

Learn frugally · Energy

Seven signals, one ELM

One frugal learner forecasts the whole Corsican electricity mix.

Pooled error against parameters stored by each model (log scale): the per-cell ELM is the lowest point; reading the whole field moves models to the right without moving them down

Learn frugally · Energy

Satellite forecasting: one cell or the whole field?

A per-cell ELM beats the same ELM reading the whole field, with a thousand times fewer parameters.

Model frugality: number of stored coefficients for a linear model and for ELMs, then the ELM diagram from the English course. Fixed and learned weights are distinguished; no energy consumption is measured.

Learn frugally · Energy

Frugal models for solar forecasting (in French)

How much model a forecast really needs before it is deployed.

Palaiseau: measured photovoltaic power, correlations over 48 lags and cumulative PCA variance. Descriptive analysis of the public dataset, distinct from the manuscript results.

Learn frugally · Energy

Direct PV forecasting: reduce the dimension (in French)

Compress the input only when the compression serves the forecast.

AI in its place

Language models help people use validated tools; the calculation, the checking and the responsibility stay with validated codes and people.

The chain: request, language model, validator, validated code, result

AI in its place · Language

Let the language model read, let the validated code compute

Execute, clarify or refuse: never guess.

Demonstrate trust

Judge against a reference, publish code, data and limits, make research findable and reusable.

The Influence Index on a doubling scale: 1 % gives −1, 2 % gives 0, 4 % gives +1, 8 % gives +2

Demonstrate trust · Research practice

Questioning how we do research: influence, integrity, excellence

With J. P. A. Ioannidis (Stanford), J. Baas and R. Boverhof (Elsevier): influence on a doubling scale, impact next to integrity, excellence as a system.

A model written by three labs, one of them an Asian partner, shared under a license and a consortium agreement; an EU industrial company integrates it into an energy management system and becomes its provider; the grid operator deploys it; the partner deploys it in its own country under local law; civil liability and contracts cover the whole chain

Demonstrate trust · Research practice

AI in research: the AI Act, a lab charter and a toolbox

Research is excluded under conditions, transfers are not: timeline, who is responsible, and ten tools for a lab.

Étude NICE, figure 5 de la version auteur : distributions des métriques pour ELM, AR, smart persistence et ARTU, réseau SIAR à trente minutes.

Demonstrate trust · Energy

When RMSE hides a difference

NICE compares a model's errors with those of persistence.

Figure 2 du manuscrit auteur public sur GitHub : champs synthétiques observé et prévu, carte gamma et carte RMSE. Les tolérances de cette version sont indiquées dans la figure.

Demonstrate trust · Energy

Beyond pointwise errors: the γ-index

A tolerance-based test carried over from radiotherapy.

Demonstrate trust · Research practice

Academic visibility: why publishing isn't enough anymore

A saturated ecosystem, and a practical toolkit for discoverable research.

Illustration conceptuelle de la visibilité numérique : un visage humain côtoie un réseau circulaire d’icônes de personnes, de communication et de données. Visuel d’accompagnement de la présentation sur l’excellence scientifique, le SEO et le GEO, sans résultat quantitatif.

Demonstrate trust · Research practice

Research excellence in the age of AI: from SEO to GEO (in French)

The RTVACOS framework: rigour, transparency, visibility, adoption.

Explore the results

How forecastable is the sun?
Solar forecastability evaluated at 68 weather stations, across ten models and 24 horizons.

Better persistence
An analytical cyclostationary reference that improves on classical persistence on the studied solar series.

Seven signals, one ELM
Seven energy outputs evaluated on six years of hourly data from Corsica.

Beyond pointwise errors
A synthetic experiment shows why small displacements can mislead a pointwise error metric.

When RMSE hides a difference
Two models with the same rounded nRMSE are distinguished by persistence-relative evaluation.

These research stories explain results from papers on which I am first author. Each page includes the question, evidence, interpretation, limitations and full collective attribution, in English and French.

Solar forecasting, made simple

Explore the ideas, methods and evidence, Prévision solaire : faire simple

Forecastability, useful baselines, error evaluation and frugal models: a guided route through our published studies, open code and ongoing research, with full attribution.