Chemists know this intuitively: mixing two fluids rarely yields a property that is the arithmetic mean of the parts. Yet many black-box models still behave as if mixtures were averages with noise.
Non-ideal mixing is the rule
Activity coefficients, association, and thickener synergy create deviations that dominate real formulations. A log-mixing rule such as Grunberg–Nissan is a better starting point than a linear blend — and still incomplete.
A viscosity sketch
- Baseline: Grunberg–Nissan log-mixing with Arrhenius temperature dependence.
- Residual: an ML correction that learns synergy the baseline misses.
- Hybrid: baseline + residual, with a confidence interval and an out-of-domain flag.
The point is not that ML is unnecessary — it is that ML should correct physics, not replace it.
Why this matters for inverse design
Inverse formulation searches a combinatorial space. If the forward model invents impossible viscosities, the search wastes experiments. Physics-informed hybrids keep candidates in a chemically plausible region before you touch the lab.
See also our Technology page and FormulaIQ.
