“My model has plateaued: is it the measurement, or the reference? Do I really have to choose between NIR, RAMAN and TERAHERTZ? Does a second technology pay for itself?”
Each technology sees what its physics lets it see. When an attribute depends on both chemistry and structure, two complementary measurements say more than one.
Data fusion assembles the signals of several technologies within one model. It pays when the second measurement brings information the first one lacks: NIR for composition and water, TERAHERTZ for porosity and coating, RAMAN for crystal form. Without that contribution, it adds cost without gain.
The proof of concept first retains the technology that answers on its own; fusion comes next, when the model residuals show what escapes it. How a proof of concept runs.
The three levels of fusion, in practice
At low level, spectra are concatenated; each block must then be weighted, or the larger one overwhelms the other. At mid level, the useful variables of each block (scores, selected bands) are extracted before being combined, for instance with multiblock PLS or SO-PLS. At high level, each technology keeps its own model and only the predictions are combined.
Hybrid model and data fusion: two distinct notions
Fusion combines several instruments. A hybrid model combines mechanistic knowledge (a mass balance, a kinetic law) with a model built on data. The two sometimes add up, but they are not the same thing.
One published example blends both logics: contents predicted by NIR or RAMAN and the compression force recorded on the press feed a single model, which predicts the dissolution profile of extended-release tablets (Galata et al., Pharmaceutics, 2019).
What it changes for the project
- Two instruments to qualify, and measurements synchronised on the same unit or the same instant.
- A model whose blocks each live and drift at their own pace: maintenance is planned at scoping. Keeping a model alive.
- A return on investment recalculated with the second measurement. What a control costs.
Where the combined model lives in production
On line, a software layer collects the stream from each instrument and the process variables, aligns them on the same instant or the same unit, runs the models and sends the result back to the control system. It serves as well to synchronise several measurement points along a line as to bring spectra and process parameters together within one model.
This is the role, for example, of ProaXesS, which deploys the models developed in Vektor Direktor. The platform is chosen together with the model, from the proof of concept onward: it must be able to run the type of combination retained.
Where it applies
- Tablets: NIR and TERAHERTZ, for dissolution and hardness. Predicting dissolution.
- Chemicals and polymers: NIR and RAMAN, for composition and crystallinity.
- Food: NIR and imaging, for composition, colour and morphology.
A measurement that has plateaued? Let us talk about what escapes it.
Forty-five minutes is enough to tell whether a second technology would bring the missing information, or whether the answer lies elsewhere.