Self-Driving Labs and Autonomous Experimentation

In November 2023 a Berkeley and Lawrence Berkeley National Laboratory system called the A-Lab reported something striking: over 17 days of continuous operation with no human intervention, it synthesised 41 compounds from a list of 58 targets, more than two new materials per day (Szymanski et al., 2023, An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature 624, 86–91). Targets came from computationally predicted stable structures; recipes were proposed by a model trained on the text of published synthesis procedures; robots weighed powders, ran furnaces, and collected X-ray diffraction patterns; automated analysis judged success, and an active-learning loop revised recipes that failed.

Then the argument started, and the argument is the more valuable half of the story.

What the loop actually closes

Strip the robotics away and a self-driving lab is a control loop over an experimental space: propose candidates from a model, execute them physically, characterise the result with an instrument, interpret that measurement into a success signal, and update the proposal model. Throughput comes from removing the human wait states between steps, so the system runs overnight and weekends and never loses a sample to a forgotten timer.

The step that carries the most risk is interpretation. Everything upstream is mechanically checkable; deciding that a diffraction pattern means "you made the target compound" is an inference, and automating it means automating a judgment that human crystallographers make with care and context.

The critique, and the correction

That is precisely where the challenge landed. Robert Palgrave and colleagues re-examined the reported diffraction data and argued that the automated phase identification had been too permissive: many of the claimed products were better explained as known, compositionally disordered versions of the predicted ordered compounds, so a large share of the 41 were neither new nor unambiguously the target (Leeman et al., 2024, Challenges in high-throughput inorganic materials prediction and autonomous synthesis, ChemRxiv). The A-Lab team responded publicly with additional data defending the syntheses.

In February 2026 Nature published an author correction (Nature 650, 2026). It records that concerns were raised about unambiguous structural identification from the diffraction data and about the novelty claim, and it revises the framing: the compounds were new to the prediction platform, not necessarily new to science. The headline claim, in other words, was substantially about the software's knowledge boundary rather than about chemistry's.

What to take from it

The dispute is not evidence that autonomous laboratories do not work. It is evidence about where their weak link sits. Automating measurement is easy and automating interpretation of measurement is not, and a closed loop that scores its own outputs will happily optimise against a lenient scorer at machine speed. The generalisation is uncomfortable and broadly applicable: throughput multiplies whatever your success criterion actually measures, including its errors.

When it breaks

• Automated characterisation sets the ceiling on every claim. Powder X-ray diffraction cannot always distinguish an ordered target from a disordered relative, and an automated matcher tuned for recall will call both a success. The instrument, not the robot, bounds what the lab can honestly assert.
• The proposal model inherits the literature's biases. Recipes mined from published procedures encode what worked and got published; failed syntheses are largely unpublished, so the model's prior is shaped by survivorship (see ML for science pitfalls).
• Novelty is a claim about a database, not about nature. "Not in our reference set" and "unknown to science" are different statements, and the gap between them is exactly what the correction addresses. Any autonomous discovery claim needs its reference set named.
• Chemistry space is narrower than the robotics suggests. A-Lab's loop covers solid-state powder synthesis with furnaces; air-sensitive, solution-phase, or single-crystal work needs different hardware and a different loop, so headline rates do not transfer across domains.
• Peer review struggles with these systems. Reviewing an autonomous lab means reviewing thousands of automated judgments, and the practical response — post-publication scrutiny of released raw data — worked here only because the data were released. That is an argument for making the release mandatory.