Concept
A driver (or key parameter) is a model input the output is materially sensitive to — one whose plausible variation moves the headline number enough to change the picture. It is the parameter where the answer’s uncertainty concentrates, as opposed to the many inputs that, however numerous, barely move the result.
Leverage, not size. What makes a parameter a driver is its leverage on the output, and leverage has two ingredients, not one:
leverage ≈ (how much the output moves per unit change in the input) × (the input's own plausible range)The first factor is the output’s sensitivity to the input — the local response, how steeply the headline number reacts. The second is how uncertain or variable the input itself is — the width of the credible range it could take. A parameter is a driver only when both are appreciable. A steep response to an input that is pinned down (its range is narrow) contributes little uncertainty; a modest response to an input that could plausibly double contributes a lot. This is the distinction that separates a driver from a parameter that is merely large or merely sensitive.
How drivers are found. Drivers are identified empirically, by varying each input across its range and measuring the output swing, then ranking the swings — the visual form of that ranking is the tornado chart, whose top bar is the dominant driver. A driver is therefore not declared by inspection but read off the model’s actual response.
Drivers concentrate. In a typical estimate a handful of parameters account for most of the output’s movement while the long tail of remaining inputs barely registers — the answer’s uncertainty is rarely spread evenly. Which parameters land at the top is set by the process’s structure: the cost structure (a capital-heavy route is driven by utilization and the capital anchor; a feedstock-heavy one by input prices), and the dominant terms of the unit-operation parameterization (a reactor’s conversion, a route’s energy intensity).
Two families of driver. Drivers split by where they come from. Physical / engineering parameters — a per-pass conversion, an energy intensity, a capacity factor — are properties of the process and its design. Market / financial parameters — an input price, a discount rate, a product price — are set outside the plant. The two families behave differently under analysis: a physical parameter is usually bounded by engineering, while a market parameter can swing on its own, which is why a market input is so often the top driver of a levelized cost.
For green ammonia’s levelized cost (~$800/t NH₃ at the continuous baseline, from the running example), a quick scan of leverage sorts the inputs. The electricity price is the top driver: at ~10 MWh/t, swinging the price across a plausible ~$30–60/MWh range (a round market band, not a sourced forecast) moves the electricity term from ~$300 to ~$600/t and the levelized cost from roughly $700 to $1,000/t — a ~±$150/t swing around the baseline. The capacity factor is comparably strong on the fixed side (0.90 → 0.45 lifts the cost by ~$400/t). By contrast, a minor vessel’s cost exponent (~0.6) barely moves the answer — low leverage — and the fixed stoichiometric H₂ requirement (~0.18 t H₂/t NH₃), though large, is not a driver at all: it is essentially fixed by chemistry and has no plausible range to move through.
Separately, to show leverage = sensitivity × range: electricity intensity (10 MWh/t) and electricity price ($40/MWh) enter the cost symmetrically — cost is their product — so their local sensitivities are equal. Yet the price out-drives the intensity, because the intensity is engineering-bounded to perhaps ±15% while the price can plausibly range ±40%. Equal sensitivity, unequal range, so unequal leverage.
And to show a driver list is model-conditional: draw the system boundary to buy hydrogen at the fence instead of making it from electricity, and the top driver flips from the power price to the hydrogen purchase price — same ammonia plant, different controlling input, purely because the boundary moved.