AI Chemistry: Keep the Experiments That Did Not Work

A laboratory experiment can miss its intended product and still leave something valuable behind: a record of what was tried, under which conditions, and what was observed. For AI-assisted discovery, that record may help define the boundary between a promising suggestion and an unproductive one.

The difficult part is preserving the meaning of the result. “Failed” can hide several different stories. A reaction produced the wrong substance. The desired product appeared only in a tiny quantity. An instrument malfunctioned. An idea was never tested. Combining those stories into one label can erase the information a future investigation needs.

A collection of successes has a particular viewpoint

A study first published on 5 May 2022 by Felix Strieth-Kalthoff and colleagues examined how experimental errors, experiment selection and incomplete reporting affect predictions of chemical reaction yields. Its case studies highlighted the importance of negative examples. The problem extends beyond having inaccurate entries: researchers’ choices about which experiments to conduct and which outcomes to report shape the available data.

Consider the question a reader should ask of a success-only archive: does it show the full set of attempts, or the selected outcomes that survived a reporting decision? This is an editorial inference from the data-selection problem. It matters when a polished AI demonstration is presented as evidence that a discovery system can explore unfamiliar territory.

Unsuccessful experiments have already supported discovery

In a Nature paper published on 4 May 2016, Paul Raccuglia and colleagues used archived laboratory notebooks containing unsuccessful attempts to make a particular family of crystalline materials: templated vanadium selenites. They combined the reaction records with chemical descriptors and trained a model to predict successful formation.

The team then tested previously untried organic building blocks in real experiments. Their model-guided approach outperformed the human strategies used for comparison. This was evidence for a defined materials-discovery task, with experimental follow-up. It supports taking unsuccessful syntheses seriously as data; it does not establish that every failed experiment will improve any model.

For general readers, the important detail is the chain of work: retained observations, a defined prediction task and a subsequent experimental check. Calling the process “AI discovery” should leave that chain visible. The notebooks and the people making and interpreting the measurements are part of the achievement.

A chemical language model needs a precise negative

A Science Advances paper published on 13 June 2025 by Alessandra Toniato and colleagues applied reinforcement learning to a specialised chemical-reaction language model. Feedback incorporated successful and unsuccessful outcomes. The researchers reported improvements over fine-tuning in low-positive-data tests and in tests using high-throughput experimental data. When the controlled dataset supplied more positive examples, fine-tuning was more effective.

The definitions matter. One controlled dataset used artificially generated incorrect product positions. A second used real high-throughput experiments, with low observed yields labelled negative using a 1% threshold chosen for the study. Those categories represent different evidence. The authors also reported training instability and sensitivity to data splits. The result concerns their models, datasets and training procedure; it gives no guarantee that adding failure examples to an ordinary chatbot conversation produces the same benefit.

Keep the observation separate from the verdict

A useful reading habit is to separate three layers in any record: what was done, what was measured and how the outcome was classified. That is a proposed way to interrogate the evidence, rather than a new scientific standard.

“Unsuitable for this purpose” is a decision with an intended use attached. A measured result may later become useful for a different question, while an unexplained failure label offers little to revisit. Ask whether the underlying observation survives alongside the classification.

• Attempt: Can a reader establish that the experiment actually took place, and identify the conditions recorded?
• Observation: What result was measured, how was it measured, and what uncertainty or recording problem remains?
• Classification: Which rule turned that observation into a positive or negative example, and for which task?
• Traceability: Can the entry be connected to its original record and any later correction?

These questions also make absence easier to interpret. An empty result field invites a follow-up about missing information. It should not silently acquire the meaning “the experiment demonstrated failure”.

A hypothetical ceramics studio shows the distinction

Imagine a small ceramics studio testing four proposed finishes. This is a fictional record-keeping example, not evidence about an actual AI system. One finish meets the studio’s visual brief. Another produces an intact surface in the wrong colour. A third test is interrupted by an equipment fault. The fourth remains an idea in a planning document.

A folder containing only the approved photograph loses much of that story. A spreadsheet labelling the other three “failed” also loses something: it merges an observed aesthetic mismatch, an interrupted test and an untested proposal.

A clearer record keeps the wrong-colour observation available, marks the interrupted test with its equipment problem and leaves the untested proposal identified as untested. The next person can decide which question each record helps answer. Perhaps the unexpected colour suits another collection; perhaps the interrupted test needs repeating.

That example illustrates a decision about information. It offers no estimate of how much data a useful model would require, and it cannot demonstrate an improvement in predictive accuracy. Its purpose is to show why a label needs an explanation.

Read an AI discovery claim through its dataset

When an article says an AI system has learned from failure, look for the meaning of failure before admiring the size of the dataset. Ask whether the negative examples came from measurements, computer-generated alternatives or a researcher’s classification rule. Then locate the comparison: what did the proposed method improve upon, and on which task?

Next, check what happened after prediction. Did the researchers assess held-out records, carry out new physical experiments, or do both? Those steps answer different questions about the system. Keep the paper’s actual evaluation attached to any description of its achievement.

MaryChuks’ guide to reading a research paper through its question, design, finding and boundary provides a practical route for this closer reading. Start with the methods and the outcome definitions when a headline makes the data sound simpler than they are.

Make unsuccessful work interpretable

For creators and business owners, the transferable idea is a record-keeping principle: retain enough context to explain why an attempt was rejected, interrupted or left unresolved. That is an editorial application of the discussion, not a claim that a laboratory result automatically transfers to marketing or product development.

AI-assisted discovery becomes more intelligible when unsuccessful work remains inspectable. The valuable resource is an observation another person can interpret, with its conditions, meaning and uncertainty intact. Before discarding an unsuccessful attempt, decide what evidence it leaves behind and what question that evidence could help answer.

Sources

• Raccuglia and colleagues, Machine-learning-assisted materials discovery using failed experiments, Nature, 4 May 2016.
• Strieth-Kalthoff and colleagues, Machine Learning for Chemical Reactivity: The Importance of Failed Experiments, Angewandte Chemie International Edition, first published 5 May 2022.
• Toniato and colleagues, Negative chemical data boosts language models in reaction outcome prediction, Science Advances, 13 June 2025.