Self-driving labs: where sampling and HPLC injection fit in the loop

A self-driving lab is an automated laboratory in which software chooses the next experiment from the results of earlier ones, hardware runs it, and instruments measure the outcome, with no human making each decision. Sampling and injection sit where the physical experiment turns into data. If the sampler cannot be scripted, cannot report what it actually did, or stops on a fault at 3 a.m., the loop stops. For liquid-phase chemistry, HPLC is often the slowest and least automation-friendly step, so the sampler and its software interface deserve as much design effort as the robot or the optimiser.

Key points

• Each landmark self-driving lab below is built around its analytical method: GC for the mobile robotic chemist, XRD for A-Lab, inline NMR for RoboChem, HPLC and UPLC-MS for several organic chemistry platforms.
• One HPLC at 10 minutes per run gives 6 results per hour. In the Merck and UBC batch platform the analysis set the pace of the whole loop.
• Analysis is also where autonomous labs have gone wrong: A-Lab's phase identification from XRD was revised in a 2026 correction.
• A closed-loop sampler needs a scriptable interface, explicit errors, fault sensors, unattended recovery and a log of executed (not requested) parameters.
• You do not need a purpose-built HPLC: one platform added an injection valve and relay triggers to a standard Agilent 1100, another used mobile robots to feed an unmodified UPLC-MS.

What is a self-driving lab?

A self-driving lab (SDL) closes the design-make-test-analyse cycle. A planner proposes conditions (Bayesian optimisation, a heuristic, or a large language model), an execution layer makes the sample (liquid handler, flow reactor, powder doser, mobile robot), an analytical layer measures it, and the result goes back to the planner. The 2024 Chemical Reviews survey by Tom et al. covers the field in depth. Funding has followed: in April 2023 the University of Toronto received a CA$200 million Canada First Research Excellence Fund grant for the Acceleration Consortium, which builds SDLs for materials and molecules.

The word "autonomous" hides a lot of plumbing. Most of the engineering effort in the papers below goes into moving samples between stations, converting a planner's suggestion into volumes a machine can dispense, and getting results back in a form the planner can read.

Landmark systems and how they analysed their samples

Three patterns stand out. First, analysis determines architecture: RoboChem could use a flow cell because NMR can be inline, while HPLC users either inject from a valve (Christensen) or move vials (Dai). Second, people usually remain somewhere in the loop: Coscientist's authors note that plates were moved manually, and Ada needed consumables replenished every seven samples. Third, several groups chose to leave analytical instruments unmodified so that they could be shared, which pushes complexity into sample preparation and transport.

Why analysis, and HPLC in particular, is often the bottleneck

Throughput. A 10-minute HPLC cycle (gradient plus re-equilibration) gives 6 results per hour, 144 per day, from one instrument. Unless sample preparation is just as slow, the HPLC sets the pace. In the Christensen et al. platform, ChemOS had the robot initiate reactions at 15-minute intervals and aliquot each one at its endpoint at the same interval, so the analysis set the pace. Ada's authors went further and replaced a slow device measurement with a faster proxy, a "pseudomobility" from conductance and absorbance, because the conventional measurement was not compatible with the timescale of autonomous optimisation.

Interpretation. An optimiser trusts whatever number the analysis returns. A-Lab's 2026 author correction followed concerns about identifying phases from diffraction alone: after manual re-analysis, the platform's conclusion held for 36 of 40 reported successes and 4 were inconclusive, and the revised paper reports 36 compounds from 57 targets. Dai et al. argue that decisions based on a single, hard-wired characterisation technique are fragile, which is why they combine UPLC-MS with NMR.

HPLC-specific friction. Samples need quenching, dilution and sometimes filtration; strong sample solvents distort peaks; very different samples in sequence expose carryover; new products need peak assignment; and chromatography data systems were designed around sequences a person sets up. Christensen et al. chose a lightweight Python framework instead of APIs between ChemOS, Chemspeed AutoSuite and Agilent ChemStation.

Practical ways around the bottleneck:

• Decouple sampling from analysis: quench aliquots into vials at the right time and inject the bank later, as described in online HPLC reaction monitoring.
• Shorten the method: short columns and steep gradients, accepting less resolution for screening rounds.
• Run a cheap screen (plate reader, UV) on everything and HPLC only on candidates.
• Add a second HPLC before adding a second robot.

What a sampler needs to work in a closed loop

A scriptable interface with documentation

Every action available from the front panel must be available as a command, with documented parameters, replies and error codes. Coscientist's documentation module let the agent correct a protocol after a failed call to the Opentrons API, which makes the point: undocumented interfaces block software agents as much as people. A plain-text protocol over a serial port is enough; controlling serial instruments from Python shows the pattern, and SiLA 2 and OPC UA LADS add discovery and typing on top.

Error reporting and sensors

The planner needs to know when a sample is bad, not only when the instrument stops. Useful signals include valve position confirmation, door state, needle collision, pressure on the sample line, missing vial, and liquid detected or not detected. Leonov et al. added colour, temperature, pH, conductivity and liquid-transfer sensors to the Chemputer and used them to detect critical hardware failures.

Recovery from faults

Overnight runs fail in boring ways: a blocked needle, a vial cap that did not pierce, a power blip. The sampler should report the fault, return to a known state (home axes, confirm valve position), and let the controller mark one sample invalid and carry on. Commands that move liquid are not safe to retry blindly, so the controller has to query state after a timeout.

Footprint, enclosure and power

Air-sensitive work may need the sampler inside a glovebox, as in Malig et al.'s Buchwald-Hartwig study, and solvent work usually means a fume hood shared with other modules. Measure the hood's usable width and depth, and the space the door and tray need (see how much bench space an HPLC takes). Low-voltage DC supplies are easier to route into enclosures than mains.

Liquid-handling range

In a loop the sampler may need to draw from a reactor, quench, dilute, add a reagent and inject. Check the volume range at both ends and how small volumes are delivered. Christensen et al.'s robot could not dispense sub-microlitre volumes accurately, so their script rounded to whole microlitres and reported the executed values back to the optimiser.

Throughput and capacity

Cycle time per injection, tray capacity and the time until a human must refill solvents, vials or tips determine how long the loop runs unattended.

Logging

Log every command, reply and sensor reading with a timestamp, plus sample genealogy: which reaction, which aliquot, which vial, which injection, which data file. Without it, a suspicious data point cannot be traced.

A requirements checklist

Where the AS-48 fits

The AS-48 is a useful worked example of these trade-offs. It appears as a virtual COM port over USB-C and uses a documented plain-text command set, so any language that opens a serial port can drive it. It has a door sensor, needle crash detection, pressure monitoring on the sample line, an angle sensor that checks every valve switch and a power-up self-check. It is 140 × 155 × 168 mm (W × D × H) and runs on 24 V DC. The HPLC configuration injects 1-100 µL; the LH configuration uses a 4 mL syringe for reactor sampling, dilution and reagent addition. Two gaps matter for a closed loop: its specification has no inject-start or contact-closure output, so starting acquisition depends on your software set-up, and there is no tray cooling. It is an engineering prototype, and values such as cycle time are design targets.