Online HPLC reaction monitoring: how automated reactor sampling works

Online HPLC reaction monitoring means a device takes a small aliquot from the reactor at set times, quenches and dilutes it, and injects it into an HPLC without anyone touching a pipette, so you get a concentration-versus-time profile for every species the column separates. The hard part is rarely the chromatography. It is sampling: a representative aliquot, chemistry stopped at a known moment, a concentration inside the detector's linear range and honest time stamps. HPLC gives one point every few minutes rather than every few seconds, but it resolves isomers, impurities and intermediates that in-situ IR or Raman cannot.

Key points

• HPLC cannot be inline; it is either at-line (sample to vial, then inject) or online (automatic draw and inject). Banking quenched aliquots in vials lets you sample faster than the HPLC can run.
• Quench and dilution decide data quality. Validate that a quenched sample does not change while it waits in the tray.
• An internal standard added to the reactor cancels aliquot-volume and dilution errors, which matter most when aliquots are 10-20 µL.
• Transfer-line dead volume costs reactor volume and adds delay: a 0.5 mm ID line holds about 196 µL per metre.
• Front-load the schedule: as a rule of thumb, most of the curvature of a kinetic profile sits in the first two half-lives.

Offline, at-line, online and inline: what is the difference?

The FDA's 2004 PAT guidance gives the definitions most process chemists use. At-line: the sample is removed, isolated and analysed close to the process. On-line: the sample is diverted from the process and may be returned. In-line: the sample is not removed at all, and the measurement can be invasive or non-invasive.

In the reaction-monitoring literature "online HPLC" usually covers both automated routes: the sampler injects directly, or it prepares vials that an autosampler injects in order. Either way, what matters is when the aliquot left the reactor and what was done to it before it reached the column.

How reactor sampling works

Sample loops in flow streams

In flow chemistry the simplest interface is a switching valve with a small loop in the reactor outlet, injecting effluent neat or after in-line dilution. Hone and co-workers used a VICI Valco internal sample injector to inject reactor effluent into the HPLC every 2 minutes while ramping the flow rate down, so each injection corresponded to a different residence time. That gave 16 kinetic profiles at 3 concentrations and 4 temperatures in under 3 hours. Loops are clean and fast, but only for homogeneous streams, and neat injections of a concentrated stream in a strong solvent can distort peaks (see how much you can inject for the solvent-strength rules).

Syringe draws from batch reactors

For batch reactors, a syringe draws through a needle or dip tube and dispenses a measured aliquot into a quench vial. In the Merck and UBC autonomous optimisation work, 10 µL of reaction mixture went into 800 µL of buffered acetonitrile quench, and 40 µL of that was injected. Most errors come from the draw path: liquid in the line is older than liquid in the flask, so purge the line before taking the aliquot.

Probe-pocket samplers

Mettler Toledo's EasySampler takes a different route: a probe in the reactor captures a fixed volume in a pocket and quenches it in situ. The EasySampler 1210 data sheet lists a 20 µL ± 10% pocket, a minimum sampling interval of 2 min 52 s, -20 °C to 140 °C at ambient pressure (up to 10 bar between 20 °C and 100 °C) and a rack of 12 × 10 mL vials. On its own it prepares vials for offline or at-line HPLC.

Quench

The quench defines the time point. Typical quenches are dilution into cold solvent, acidic or buffered diluent that kills a base or catalyst, or a scavenger for a reactive reagent. Malig et al. chose a copper-catalysed azide-alkyne cycloaddition as one test case precisely because it is difficult to quench due to very high catalyst activity. Always test quench stability: inject one quenched sample immediately and again after the longest wait your sequence allows. Christensen et al. switched from offline LC to online HPLC after uncovering analytical sample instability.

Filtration and slurries

Solids block lines and valves and make aliquots unrepresentative. You can either sample the solution phase only, through a filter at the probe (Mettler Toledo's EasyFrit uses a 10 µm pore size), or take the whole slurry through wide-bore tubing and filter or dissolve afterwards. Telescope Innovations, whose DirectInject-LC is now distributed by Mettler Toledo, notes that loop circulation and capillary extraction work well for uniform liquids but poorly for solids; their system samples 20 µL as often as every 3 minutes and dilutes up to 5000-fold before injection.

Dead volume and timing

Every transfer line holds old liquid. Its volume is:

V_line = π × (d/2)² × L

Worked example: a 0.5 mm ID PTFE line, 600 mm long, holds π × 0.25² × 600 = 118 mm³, so 118 µL. A common rule of thumb is to purge about three line volumes (around 350 µL here) before taking the aliquot. If that purge goes to waste, 18 samples remove 18 × (350 + 20) µL = 6.7 mL from the reactor, 13% of a 50 mL batch. Return the purge to the reactor, or use narrower and shorter tubing, and the loss drops to the aliquots themselves (0.36 mL).

The time stamp belongs to the moment the aliquot is isolated and quenched, not to the injection. For flow reactors, convert injection time back to residence time using the reactor volume and the flow rate history, as Hone et al. did for their ramped experiments.

Published systems and commercial tools

Agilent's application note on a Vapourtec R-Series flow reactor is a good template for fast chemistry. The Online Sample Manager drew samples every 30 seconds to 2 minutes, diluted them 1:100 to 800 µL, banked them in the tray and injected them in order, with about 7 minutes between injections. Sampling and analysis were decoupled, so the time resolution was set by the sampler, not by the gradient. Chemspeed also offers sampling from its parallel reactors to online UPLC-MS for closed-loop optimisation, and the same pattern sits at the centre of most self-driving lab set-ups.

Time stamps, peak tracking and data handling

Log these for every sample, from one clock, in ISO 8601 with a time zone:

• Reaction time zero (last reagent or catalyst added) and the quench time of each aliquot.
• The executed aliquot and diluent volumes, not the requested ones. Christensen et al. had to round sub-microlitre requests to whole microlitres and fed the executed values back to the optimiser.
• Vial position, injection time and the data file name.

Put an inert internal standard in the reactor, as Christensen et al. did with 1,3,5-trimethoxybenzene. Assay yields then come from peak area ratios, and errors in aliquot volume, dilution and injection volume cancel. For peak tracking across 20-100 chromatograms, identify peaks by retention time relative to the internal standard, confirm with DAD spectra or MS, and keep unassigned peaks in the table as area percent: the peak that appears at 10 minutes and is gone by 60 is often the intermediate you are looking for.

Synchronisation is the last piece. Christensen et al. used relay switches to start chromatography and detection when their robot injected. If your sampler has no trigger output, the controlling script has to start the acquisition through the data system's own interface, or the HPLC method has to be started by the instrument that injects. Writing that script is covered in controlling serial lab instruments from Python.

Worked example: a sampling schedule and dilution

A batch reaction starts at 0.20 M substrate (MW 250 g/mol, so 50 mg/mL). A pilot run suggests roughly first-order decay with a half-life of about 20 min, so k = ln 2 / 20 = 0.035 min⁻¹. The HPLC method takes 8 min including re-equilibration. You want 3 hours of data.

Schedule. Space samples evenly in conversion rather than in time:

That is 18 samples. With direct injection, samples could be no closer than the 8 min HPLC cycle, so the 2, 4 and 6 min points would be impossible. If you quench into vials and inject them in order as soon as the HPLC is free, the queue is at most 42 min behind (the 30 min sample is injected at 72 min) and the last result arrives 8 min after the last sample. The quench therefore has to hold for at least 45 minutes in the tray.

Dilution. Suppose your calibration covers 0.1-2 µg of substrate on column and you aim for about 1 µg at t = 0.

DF     = (V_aliquot + V_diluent) / V_aliquot
c_vial = c_reactor / DF
m_col  = c_vial × V_inj

Option A, one step: 20 µL into 1480 µL gives DF = 75 and c_vial = 0.67 mg/mL; a 1.5 µL injection puts 1.0 µg on column. Option B, two steps: 50 µL into 950 µL (20×), then 100 µL of that into 900 µL (10×), gives DF = 200 and c_vial = 0.25 mg/mL; a 4 µL injection gives 1.0 µg. Option A uses one vial but needs a precise small injection; option B uses two vials and is easier on the injector. Both keep the vial at 1.5 mL, a sensible fill for a 2 mL vial.

Check the aliquot step gravimetrically: a 0.5 µL error is 2.5% of a 20 µL aliquot but 6.7% of 7.5 µL. An internal standard in the reactor removes most of this error.

Where the AS-48 fits

The AS-48 is built for the at-line version of this workflow. In its LH configuration a 4 mL syringe samples a reactor, dilutes, adds reagents and dispenses into 48 × 2 mL vials; in the HPLC configuration it injects 1-100 µL into an HPLC flow path. You switch configurations with a syringe cartridge and a tube kit. Its specification has no inject-start or contact-closure output, so starting the HPLC acquisition in step with an injection depends on your set-up, typically a script that drives the sampler over its serial port and starts the data system through its own interface.

Choosing a set-up

• Fast chemistry (half-life under 10 min): sample and quench into vials, bank them, inject later. Direct injection cannot keep up.
• Homogeneous flow stream: a valve loop is the simplest and most robust interface.
• Slurries or crystallisations: filtered probes or wide-bore whole-slurry sampling, never a circulating loop.
• Air-sensitive chemistry: keep sampling inside the inert enclosure (Malig et al. used a glovebox).
• Whatever the hardware, validate the quench, add an internal standard and log executed volumes and quench times. Early and late samples differ a lot in composition, so measure autosampler carryover, and size the loop with full-loop vs partial-loop injection in mind.