In a university catalysis lab, a graduate student was running a routine reproducibility check. She had been assigned to replicate a set of ligand screening experiments that had produced clean, publishable selectivity curves six months earlier. But the new runs looked nothing like the old ones. Yields were off. Ratios of product isomers had flipped. The student spent weeks checking instrument calibration, rerunning standards, and swapping columns. Nothing fixed it. Only when she retrieved the original solvent bottle from a waste container did the explanation emerge: the solvent batch used in the original experiments was not the same as the fresh bottle on the shelf. That single unindexed lot had reshuffled 14 of 20 catalytic selectivity curves.

The Solvent Batch That Broke the Curves

The effect was not subtle. When the student re-ran the original reactions with the archived solvent, the selectivity curves snapped back into alignment. Fourteen out of twenty catalyst–substrate combinations showed statistically significant shifts in enantiomeric or regioselectivity when the solvent lot changed. Some curves moved by more than 20 percentage points. The team initially suspected instrument drift or human error, but the pattern was consistent: the old bottle produced one set of results, the new bottle a different one.

The solvent in question was a common polar aprotic solvent, purchased from a major chemical supplier. Its certificate of analysis listed purity above 99.8%, with water content within specification. No batch-specific impurity was flagged. Yet comparative gas chromatography–mass spectrometry revealed a distinct peak in the original bottle that was absent from the fresh lot. The impurity was not identified on the supplier's spec sheet, and its concentration was below the level that would trigger a purity downgrade.

Trace metal contamination is one plausible explanation. Many catalytic reactions are exquisitely sensitive to parts-per-million levels of metals such as iron, nickel, or chromium, which can coordinate with ligands or alter oxidation states. The impurity peak in the GC-MS did not match common metal complexes, but it could have been an organic byproduct from the solvent manufacturing process that acted as a ligand or inhibitor. The authors of the eventual reproducibility report noted that the impurity's identity remains unknown, and that the effect was only visible because the screening panel was large enough to reveal a systematic shift.

The incident echoes a broader pattern in experimental chemistry: reagents assumed to be interchangeable are not. A similar case with a Ziegler–Natta catalyst batch showed how a single contaminated lot can distort polymer stereoregularity curves. In both instances, the variable was not a deliberate experimental parameter but a logistical detail—a lot number that was never recorded.

How a Routine Reagent Became a Hidden Variable

The solvent was ordered from a supplier that stocks hundreds of thousands of liters per year. The lab's purchasing system logged the catalog number and quantity, but not the lot number. The original experimenter, a postdoc who had since left the group, had not written the lot number in the lab notebook. This is standard practice in many labs: lot numbers are considered irrelevant because the chemical is assumed to be identical across batches.

But chemical manufacturing is not perfectly reproducible. Different lots can have different impurity profiles due to variations in feedstock, distillation conditions, or storage time. Suppliers test for a handful of specified impurities, but they do not test for every possible contaminant. A batch that passes all specs may still contain an unknown compound that affects a sensitive catalytic system. The probability of such an effect is low, but when it occurs, it can derail an entire screening campaign.

The lab's own quality control did not include routine GC-MS of every solvent bottle—that would be prohibitively expensive for a high-throughput lab that uses dozens of solvents per week. So the hidden variable remained hidden until a serendipitous audit.

The effect was only detectable because the screening data were collected in a systematic, high-throughput format. The original 20 selectivity curves were generated on an automated platform that ran the same set of reactions under identical conditions. This allowed the graduate student to compare the old and new datasets point by point. In a traditional one-off reaction, a batch effect might be mistaken for random noise or attributed to a different cause.

The Audit That Uncovered the Artifact

The reproducibility check was part of a larger initiative by the lab's principal investigator to validate published results before building on them. The graduate student was tasked with replicating the selectivity curves as a baseline for a new catalyst design. She followed the original protocol exactly, using the same ligand stock solutions, the same substrate concentrations, and the same temperature profile. But the yields were consistently lower, and the enantiomeric excess values were scattered.

After exhausting instrument diagnostics, she began comparing the reagents. The substrate and ligand bottles were from the same lots as the original. The solvent, however, came from a new bottle that had been opened a week earlier. She checked the original lab notebook: no lot number. She checked the purchasing records: the original order had been fulfilled two years earlier, and the lot number was not recorded in the system. The only way to get the original solvent was to find the old bottle.

Fortunately, the lab had a waste storage area where empty chemical bottles were kept for a few months before disposal. The graduate student searched through dozens of containers and found the original bottle, still with a few milliliters of solvent at the bottom. She carefully transferred the residue and ran the reactions again. The selectivity curves matched the original data within experimental error. She then ordered a fresh bottle of the same solvent from the same supplier but with a different lot number. That bottle produced the shifted curves. The batch effect was confirmed.

Comparative GC-MS of the old and new solvent samples revealed a small peak at a retention time of roughly 12.3 minutes in the original lot that was absent in the new one. The peak's area corresponded to an impurity concentration of about 0.02%, well below the 0.2% threshold that would trigger a purity warning. The impurity was not identified, but its presence correlated perfectly with the selectivity shift. The team concluded that the impurity acted as a modulator, possibly by coordinating to the metal center or by altering the solvent's polarity at the reaction interface.

The incident was written up as a case study in reproducibility, and the lab changed its protocols: now every solvent bottle is assigned a barcode, and the lot number is recorded in the electronic lab notebook before any reaction is run. But the broader implications extend far beyond one lab.

What This Means for High-Throughput Experimentation

High-throughput screening (HTS) is a cornerstone of modern catalysis research. Automated platforms can run hundreds or thousands of reactions per day, generating massive datasets that feed into machine-learning models for catalyst optimization. The assumption is that all variables except the intended ones are held constant. But if a hidden variable like solvent lot changes halfway through a screening campaign, the dataset becomes inconsistent.

Consider a typical HTS campaign: a lab orders a 20-liter drum of solvent at the start of a project. After several weeks, the drum runs out, and a new drum is ordered from the same supplier. If the new drum has a different impurity profile, all reactions run after the switch will be on a different baseline. If the lab does not record the lot number, the break point is invisible. The resulting dataset will contain a systematic shift that could be misinterpreted as a real catalytic trend.

Machine-learning models trained on such datasets will learn the impurity effect as if it were a feature of the catalyst or substrate. The model's predictions will then fail to generalize to new batches of solvent. This is a form of dataset shift that is particularly insidious because it is not obvious from the data alone. Only by tracking lot numbers and, ideally, archiving small samples of each reagent lot can the shift be detected and corrected.

Some labs have begun barcoding every reagent container and using automated inventory systems that record lot numbers at the point of use. But this practice is far from universal. In a 2024 survey of academic chemistry labs conducted by the Reproducibility Project: Chemistry, fewer than one in three reported routinely recording lot numbers for solvents. The cost of implementing such a system is low—barcode stickers and a simple database—but the cultural change is slow.

The problem is amplified in collaborative projects where reagents are shared across institutions. A solvent batch used in one lab may not be available in another, making replication difficult. Shared equipment and reagent initiatives have improved access, but without standardizing lot tracking, they can introduce new sources of variation.

Computational Workflows Miss the Physical World

The solvent batch incident also exposes a gap between computational chemistry and experimental reality. Density functional theory (DFT) calculations and molecular dynamics simulations typically treat the solvent as a continuous dielectric medium or as an ensemble of ideal molecules. Impurities are not included because they are not known. The predicted selectivity rankings from these simulations often disagree with experimental results, and the discrepancy is usually attributed to inadequacies in the theoretical model.

In this case, the DFT calculations had predicted a different ordering of selectivity for the catalyst series than what was observed experimentally. The team had spent months trying to improve their computational model by testing different functionals and solvation models, but the mismatch persisted. Only when the batch effect was discovered did they realize that the experimental data they were trying to match were not a pure expression of the catalyst's intrinsic behavior. The impurity was biasing the results in a way that the theory could never capture.

When the experiments were rerun with the original solvent batch, the selectivity rankings shifted to align with the DFT predictions. The computational model had been correct all along; the experimental data had been contaminated. This is a sobering lesson for the growing field of computational catalysis, where researchers often rely on published experimental benchmarks to validate their methods. If those benchmarks are themselves artifacts of unrecorded reagent lots, the validation is meaningless.

The incident underscores the need for coupling metadata with computational inputs. A DFT calculation that uses a solvent model should ideally have access to the actual impurity profile of the solvent used in the experiment. This is not currently standard practice. Most computational studies simply cite the solvent name and purity grade, assuming that the supplier's specification is sufficient. But as this case shows, the specification can miss the critical detail.

Some computational groups have started to collaborate with experimentalists to archive small aliquots of the actual solvent used, so that if an anomaly appears, the impurity can be characterized and incorporated into the model. This is a step in the right direction, but it requires a level of coordination that is rare in the current funding and publication culture.

Lessons for Reproducibility Infrastructure

The solvent batch incident is a microcosm of a larger reproducibility challenge. The problem is not that scientists are careless—it is that the infrastructure for tracking experimental details is not designed to catch subtle batch effects. Electronic lab notebooks (ELNs) are widely used, but they often lack a dedicated field for supplier lot numbers. A simple dropdown menu or barcode scanner could solve this, but most ELN vendors have not prioritized it.

Centralized reagent tracking systems, such as those used in pharmaceutical companies, are rare in academic labs. These systems assign a unique identifier to every container and log every dispense event, creating a complete chain of custody. The cost of implementing such a system in a university setting is modest, but the return on investment is difficult to quantify until a batch effect is discovered. After this incident, the lab's PI estimated that the wasted time and materials from the unreproducible results amounted to roughly six months of a graduate student's effort, plus several thousand dollars in consumables.

Journals could also play a role. Some chemistry journals now require authors to report the lot numbers of key reagents, but enforcement is inconsistent. A stronger requirement—such as archiving a sample of each reagent lot in a public repository—would be a significant step. Funding agencies could mandate such archiving as a condition of grant support, similar to data management plans. The cost per sample is negligible, and the benefit to the community is large.

But there are counter-arguments. Requiring lot number reporting for every reagent could be burdensome, especially for labs that use hundreds of different chemicals. The vast majority of lot-to-lot variations are harmless, and the effort to track them all might outweigh the benefit. A more targeted approach would be to require lot tracking only for reagents that are known to be variable or that are used in sensitive catalytic reactions. However, this requires prior knowledge of which reagents are sensitive—knowledge that is often lacking until an incident occurs.

Ultimately, the choice between comprehensive tracking and targeted tracking depends on the lab's resources and risk tolerance. A high-throughput lab that screens hundreds of catalysts per week may find the cost of barcoding every solvent bottle trivial compared to the cost of a false lead. A small lab running a handful of reactions may not. The reproducibility crisis in science is often framed as a problem of fraud, p-hacking, or sloppy statistics. But this case illustrates a more mundane and perhaps more common cause: the failure to record a simple piece of metadata. A solvent lot number, a catalyst batch number, a column serial number—these are the unglamorous details that can make or break a study. The infrastructure to capture them exists, but it is not yet standard. Whether the benefits of universal tracking justify the costs remains an open question, one that each lab must answer for itself based on its own experience and priorities.