Understand before you compute: the approach of Cyril Voyant

Medicine, energy, language: very different questions, one method. People and their question at the center; a model we understand at the core; every tool, artificial intelligence included, around it. Complexity is added only when it brings a measurable gain.

Cyril Voyant · Research Director, Mines Paris – PSL, Centre O.I.E.

In brief

Map of the work: four columns for the steps of the approach (direct model; hybrid, model plus learning; frugal learning; heavy learning and language models) and three rows for the fields (medicine, energy, language), each published or open work written in its cell
Map of the work: four columns for the steps of the approach (direct model; hybrid, model plus learning; frugal learning; heavy learning and language models) and three rows for the fields (medicine, energy, language), each published or open work written in its cell

One method, every field

Cyril Voyant works on the dose of a radiotherapy treatment, the response of thyroid cancer to radioactive iodine, the output of a solar plant, the electricity mix of an island, the load of a road tunnel, and requests written in plain language. Medicine, energy and language are treated the same way. The domain changes; the method does not.

1. Start with the direct approach

When the link is known, compute it. Radiobiological dose equivalence follows an explicit model, checked on 4,438 treatment schedules (LQL-Equiv). When no measurements exist, the electricity demand of a tunnel is simulated from what is known about it (Tunnel Load Simulator). Before forecasting the sun, describe how variable and how forecastable a site is (sCV and F).

2. When the model falls short, hybridize

This is the route Cyril Voyant prefers. When an explicit model meets the need well enough, no learning is added without reason. When it falls short, it is completed, not replaced: learning informed by physics, medicine or mathematics estimates only what the model does not capture.

3. When relations are not known well enough, learn, only as far as it pays

Persistence, then structured references (BLEND, CLIPER, ARTU), statistical models, frugal learning machines (ELM), and only then deep networks or foundation models. Each step is kept only if it beats the previous one under the same protocol.

4. Artificial intelligence, in its place

Machine learning, deep learning and language models are valuable and used every day. They remain tools around an understood, validated core. In WebSemantic, a language model reads the request, a validator asks whenever an input is missing, and a validated code computes every number.

5. Trust is demonstrated

Every result is judged against a reference with metrics that match the question (NICE, the temporal γ-index), and published with its code, data and limits, archived with a DOI. Research must be found, trusted and reused: this is the point of the RTVACOS framework on research excellence. The same questioning applies to research itself: with J. P. A. Ioannidis (Stanford University), J. Baas and R. Boverhof (Elsevier), institutions are mapped by impact next to research-integrity proxies (Royal Society Open Science, 2026), and influence is read against an explicit 2 % baseline rather than by volume (questioning how we do research).

The five principles

  1. People at the center: forecasts and calculations serve a human decision.
  2. The model at the core: understand first; compute directly; hybridize when the model falls short.
  3. Complexity only if it pays: no a priori on the method; start simple.
  4. Every tool, at the periphery: AI included, often in hybrid form.
  5. Trust is demonstrated: evaluation, open code, data and limits.

Go further