A freshly picked liquid class is a hypothesis, not an answer. It says "this liquid, on this instrument, with this tip, probably behaves like this," and probably is doing a lot of work in that sentence. Optimization is the process of replacing probably with measured. It is the least glamorous part of automated liquid handling and the part that most decides whether your results are trustworthy, so it is worth understanding before you touch a single parameter.
The word optimization sounds like it belongs to mathematicians running solvers over a search space. In the lab it is far more modest and far more human: you make one small change, you watch what it does, and you keep the changes that help. Most experienced practitioners do not run a formal design of experiments. They start from a class that is close, adjust incrementally, and measure until the transfer is both repeatable and correct. That loop is the whole discipline. This guide is about understanding the loop before you run it.
Why a default class is never the finished class
Every liquid handling platform ships with default classes, and every one of them was built for water or something close to it, on reference hardware, at a temperature that is not your temperature. That default is a genuinely useful starting point, which is why the fastest new class always begins by cloning the closest existing one rather than starting from a blank form. But the moment your liquid stops being water, the defaults stop being right.
The reasons are physical, not arbitrary. A liquid class is a translation of how a liquid behaves into numbers an instrument can follow, and behavior is set by properties the default never saw.
- Viscosity: thick liquids resist the plunger, so suction that outruns the fluid pulls in air and dispenses that outrun it leave liquid behind. Glycerol at ten percent of the water flow rate is a common starting point, not an exaggeration.
- Volatility: solvents like ethanol and acetone evaporate off the tip and drip during transport, so a class tuned for water will lose volume before it ever reaches the well.
- Surface tension and cohesion: some liquids cling to the tip and refuse to let go, others form droplets that fall on their own. The air column and the exit speed have to be tuned to the liquid, not assumed.
- Your hardware and labware: the same intent needs different numbers on a Hamilton than on a Tecan or an Opentrons, and tip geometry interacts strongly with low volumes. A class is only valid for the tip it was built on.
None of this means the default is wrong. It means the default is a first draft, and optimization is the editing.
The two things you are actually optimizing
Beginners often chase a single number, "am I delivering 50 microliters," and get frustrated when it wobbles. The fix is to separate two ideas that feel like one.
- Precision is repeatability: how tightly your deliveries cluster together, run after run. It says nothing about whether the cluster is in the right place. A class that reliably delivers 47 microliters every time is precise.
- Trueness is accuracy: how close the average delivery sits to the volume you asked for. Our example class is precise but not true, because its average is off by three microliters.
The order matters, and it is the single most important thing to learn early: optimize precision first, trueness second. Precision comes from the mechanical parameters, the flow rates, the settling times, the air gaps, the exit speeds, everything that governs how cleanly the liquid enters and leaves the tip. Trueness comes at the end from the correction curve, a simple mapping that tells the instrument to command a little more or a little less so the average lands on target.
You cannot correct your way out of a wobble. If the spread is wide, a correction curve just shifts a blurry cluster onto the target, and it is still blurry. Tighten the cluster with the mechanical parameters until the spread is acceptable, and only then nudge the whole cluster onto the bullseye. Reverse the order and you will chase your own tail for an afternoon.
The optimization loop, once around
Here is the shape of the whole thing, so the rest of the path has a skeleton to hang on. Every step gets its own detailed treatment later, but this is the loop you are learning to run.
- Start close. Clone the validated class for the most similar liquid you have, so your adjustments are small and targeted rather than sweeping.
- Test a real transfer. Run one step that mimics your actual method and watch it happen. Visual inspection catches most problems before any balance does.
- Change one parameter at a time. Adjust a single knob, re-run, and see what moved. It is slower per iteration and much faster overall, because you learn what each knob does for this liquid instead of stumbling onto a lucky combination you cannot reproduce.
- Measure. Once it looks clean, weigh replicates or read a dye to get real delivered volumes and a real spread.
- Correct. With precision in hand, set the correction curve across your working range so the average matches target at each volume you care about.
- Verify and record. Confirm the class holds up, then write down what you did and why, so the next person does not start from a blank form either.
Knowing when to stop
Optimization has no natural end, so you have to give it one. The stopping point is not perfection; it is fitness for purpose. Decide up front what precision and trueness your assay actually needs, because a normalization step and a qualitative spot check do not demand the same tolerance, and tuning past your requirement is time you are not getting back. When the measured spread and offset sit inside the tolerance your method needs, across the volumes you actually run, the class is done. Stop, record it, and move on.
A default class is a first draft in someone else's handwriting. Optimization is you editing it into the truth, one measured change at a time, precision before trueness.