mathematics//dynamical systems//model and representation//grey-box model

A grey-box model is a dynamical model whose equations come from physics and whose parameters are fitted to measured data, and it is the workhorse of industrial modelling: the thermal model of a motor, the thrust and inertia model of a drone, the heat balance of a furnace. It sits between two pure cases. A **white-box model** is first principles with every parameter measured directly (a pendulum, a tank with a known orifice); it needs almost no data and extrapolates well as long as the physics is right. A **black-box model** is data alone (an ARX model, gradient boosting, a neural network trained on the history of a production line); it needs plenty of data covering the whole operating range and does something unpredictable outside it.

A grey-box model is a dynamical model whose equations come from physics and whose parameters are fitted to measured data, and it is the workhorse of industrial modelling: the thermal model of a motor, the thrust and inertia model of a drone, the heat balance of a furnace. It sits between two pure cases. A white-box model is first principles with every parameter measured directly (a pendulum, a tank with a known orifice); it needs almost no data and extrapolates well as long as the physics is right. A black-box model is data alone (an ARX model, gradient boosting, a neural network trained on the history of a production line); it needs plenty of data covering the whole operating range and does something unpredictable outside it.

The grey box takes the best of each. Physics gives the form, which is the hard part to learn from data: that thrust grows with the square of rotor speed, that a housing heats through one capacitance and cools through one conductance. Data give the numbers, which are the hard part to measure: the thrust coefficient of this propeller, the effective heat capacity of this motor with its mounting. The fit needs moderate data, provided the tests excited the system enough to separate the parameters (persistent excitation).

Physics for the form, data for the numbers, is the default choice in industry.

Out of the training distribution a grey box degrades gently because the physics still holds; a black box can do anything, and a model that fits test data slightly better is no evidence that it will.

A modern variant adds a learned correction to the physics, a hybrid residual model:

x˙=fphys(x,u;θ)+gϕ(x,u).\dot x=f_{\text{phys}}(x,u;\theta)+g_{\phi}(x,u).x˙=fphys​(x,u;θ)+gϕ​(x,u).

θ\thetaθ are the fitted physical parameters (mass, thrust coefficient) and gϕg_\phigϕ​ a small data model that learns only what the physics misses (rotor aerodynamics, friction). If the physics is good the residual is small, needs few data and, when it is wrong outside the data, it is wrong by little.

Data teach the physics of the data given. A drone model trained on calm days has never seen wind, and a neural network being a universal approximator says no more than that a giant table is one: the difference lies in what happens on inputs nobody recorded (out-of-distribution). Putting the physical equations inside a network's loss (physics-informed learning) is mostly research.

Choosing a box is choosing a cost. The white box costs engineering time and fails when an effect was left out; the black box costs data and fails silently off its range; the grey box costs some of each. How the fit is done and checked is system identification, and the broader trade between modelling and learning is model versus data.