Fund the measurement chain before the excitation chain. A water-splitting system is limited by its gas loop and its detector, so those come first; the light source can be upgraded later without invalidating earlier data, provided spectrum and irradiance were recorded. A 65 mL loop with a leak rate of ≤0.1 μmol/h (O₂) beats a brighter lamp on a leaking rig every time.
Every subsystem adds noise to the same measurement, and the noisiest one decides what you can report. A detector that resolves 0.05 μmol of H₂ is wasted on a loop that leaks 0.1 μmol/h of O₂ across the same period, because the leak sets the floor and the detector only describes that floor precisely.
That is the budget argument in one line: spend where the noise is, not where the catalog looks impressive. Light is the most visible component and the least often limiting one — a 300 W xenon source delivering 50 W of total optical power is more excitation than most slurry systems can use without a filter.
The practical move is to measure your current floor. Run a blank, compute the spread in μmol/h, and compare it with the rate you hope to detect. The gap tells you which subsystem to fund next.
Order them by what their failure invalidates, not by what they cost. The first three decide whether any number you produce is meaningful; the rest decide how convenient the work becomes.
| Order | Subsystem | Can it wait or be upgraded later? | What its failure invalidates |
|---|---|---|---|
| 1 | Gas loop, vessel and seals | No — it defines the blank | All rates; a leak adds or removes product silently |
| 2 | Quantification: GC or trace gas analyzer | No — resolution sets the floor | Comparisons near the detection limit |
| 3 | Light source | Partly — power can rise, spectrum cannot change cheaply | Quantum yield and any cross-paper comparison |
| 4 | Reactor and window geometry | Partly — a second reactor can be added | Areal rates and path-length normalization |
| 5 | Thermal control | Yes, to a point | Arrhenius comparisons and photothermal claims |
| 6 | Logging and sampling accessories | Yes, fully | Only transients shorter than the log interval |
Read the third column as a risk statement. Anything marked "no" has to be right at purchase, because replacing it later means repeating every experiment done before the change. Anything marked "yes" can be added without touching the validity of earlier data, provided you recorded the conditions you ran under.
The gas loop and the detector cannot be bought cheaply and fixed later: the first sets the blank, the second sets the resolution. A leak is not a nuisance, it is a bias — hydrogen leaves faster than it should, oxygen enters from the room, and both move the apparent rate in a direction you cannot correct afterward.
Line volume belongs to the same decision. A compact 65 mL loop reaches equilibrium quickly and dilutes the product little; a large dead volume adds a time constant and a dilution factor, and neither is recovered by a better detector. When a supplier quotes a leak rate, ask which volume it was measured on.
Quantification deserves equal weight. An analyzer resolving 0.05 μmol of H₂ and 0.0005 μmol of CH₄ and CO describes rates that a coarse detector cannot, and where the expected product is small that resolution is the experiment.
Logging, an autosampler, a second reactor and extra detectors can all arrive after the first paper without invalidating it. The rule is that the addition must not change the optical or thermal conditions of the runs already published.
What cannot be added later is spectral coverage. The channel count of an LED head is fixed at purchase and the lamp type fixes the spectrum; if the chemistry needs a band you did not buy, no accessory will produce it.
The cheapest components sit at the boundary between the reactor and the measurement, which is exactly where a small saving becomes a large error. Septa, tubing, valve seats and fittings are the usual suspects.
A septum that leaks only when warm passes a cold leak test and fails on the third illuminated run. Tubing that adsorbs trace products delays the peak and flattens it, which reads as a slower reaction rather than as a plumbing problem. A valve with an elastomer seat absorbs the species you are counting and releases it into the next run.
None of these announce themselves. They appear as scatter, as a blank that drifts upward, or as a rate that falls whenever a new batch of consumables is installed — and the natural response is to blame the catalyst.
Yes, if its spectrum matches the absorption of your material and you record irradiance at the reactor plane, the working distance and the filter used. A documented source is comparable; an undocumented one turns every later comparison into a guess. The lamp is the easiest subsystem to upgrade later.
Buy the one that matches your expected rate. At low production rates the analyzer wins, because resolution near the floor is the whole measurement; for product identification and multi-component work the GC wins. If both matter, take the analyzer first, since a rate you cannot resolve cannot be identified either.
The one that currently sets your noise floor. Run a blank, look at the spread in μmol/h, and compare it with the smallest difference you want to claim. Upgrade whichever subsystem that spread points to; upgrading the light source first usually changes the rate without improving confidence in it.
