The Most Valuable Experimental Data May Be the Data You Never Record

As AI becomes more capable and robotic experimentation platforms become more common, I keep returning to a question: how can experimental scientists make the most of these tools? If AI can help design experiments, execute protocols, and analyze results, how can the intuition built over years at the bench be brought into that workflow to make experimentation better?

So I took one of my own papers, Choi et al. in ACS Applied Materials & Interfaces, on p-type electron-transfer doping in mixed Sn-Pb perovskites, and dissected it from start to finish.

The short answer: following the structure of the paper, the line between what AI handles instantly and what still depends on an experimenter’s intuition turned out to be remarkably clear.

By the end of this post, I hope you’ll have a more concrete way to think about where experimental expertise still matters, and how the tacit knowledge built at the bench could become an essential part of AI-driven and autonomous experimentation.

Here’s that process, step by step.

In This Post

1. When a Theoretically Sound Hypothesis Fails
2. The Reasons Protocols Don’t Tell You
3. Hidden Experimental Variables
4. When You Can’t Take Data at Face Value
5. The Things That Never Get Written Down
Where AI’s Role Ends

1. Introduction: Where a “Theoretically Sound Hypothesis” Broke Down Experimentally

This paper sequentially dopes MAPb0.5Sn0.5I3 films with two molecular dopants, F4TCNQ and Mo(tfd-COCF3)3, via sequential deposition.

The logic the authors lay out in the Introduction and Figure 2 is straightforward:

Perovskite HOMO (valence band edge): -4.95 eV
F4TCNQ LUMO: -5.24 eV — relatively shallow
Mo(tfd-COCF3)3 LUMO: -5.64 eV — deeper, a stronger oxidant

Per standard energy-level-alignment theory, the dopant with the deeper LUMO — the stronger oxidant, Mo(tfd-COCF3)3 — should pull electrons from the perovskite HOMO more efficiently, and so should show higher doping efficiency.

The paper itself sets up exactly this hypothesis. But the actual measured doping efficiencies, defined as the number of carriers generated by doping divided by the number of dopant molecules incorporated into the film, came out essentially identical:

F4TCNQ: 0.031(3)%
Mo(tfd-COCF3)3: 0.024(3)%

The two values overlap within error.

If you ask an AI, “Why did this hypothesis fail?”, it can accurately reconstruct the explanations the paper itself already offers: molecular size, steric constraints, Lewis basicity, and differences in how each molecule interacts with the film surface. The problem comes next.

Ask, “So what molecule should we screen for next to avoid this trap?”, and the AI can pull a list of candidates with suitable LUMO values from the literature. But that first-pass screening still does not tell you reliably whether the molecule will aggregate or spread evenly on the actual film surface. Ultimately, that behavior has to be validated experimentally by making the solution and seeing how it behaves during deposition, which is exactly what shows up later in the paper as the solubility gap between F4TCNQ (~0.5 mg/mL) and the Mo compound (~30 mg/mL).

AI can quickly build first-order hypotheses from concepts like energy-level alignment. Experimental intuition becomes especially valuable when those hypotheses are shaped by second-order effects such as how a molecule actually behaves at the film surface.

2. Experimental Section: Numbers With No Reasons

Reading the Experimental Section, the numbers are precise, but why this number? is almost never explained. Feed this section to an AI and it reproduces the protocol text well. But when the rationale is not written in the paper, it can at best infer one; it cannot know whether that inferred rationale was actually what guided the experimenter.

Why Sequential Deposition Was Chosen

The paper states that it demonstrates p-type doping via a sequential approach, depositing the perovskite film first and the dopant afterward, specifically to keep the doping step from disturbing perovskite film quality.

Compared with simultaneous deposition — as in Euvrard et al.’s F4TCNQ work, where perovskite and dopant are deposited together — this choice comes from hands-on experience. The solvent in a dopant solution applied mid-spin-coat could redissolve already-formed crystals or disrupt the morphology. The tradeoff is that the dopant then stays confined near the surface rather than penetrating deeply, as confirmed later by the 3D TOF-SIMS results in Figure 6.

And unless this tradeoff is represented in the literature, prior experimental records, or the prompt itself, there is no guarantee that an AI system would identify it at the experimental-design stage.

The “Shelf Life” of Precursor and Dopant Solutions

Perovskite precursor solutions: filtered and spin-coated without an antisolvent within 30 minutes of preparation
Dopant solutions (100 μL): deposited onto each film within 30 minutes of preparation. That 30-minute window carries meaning that is easy to miss unless you understand how the solutions change through oxidation over time.

When an AI auto-generates a protocol or ports it to a robotic platform, this kind of time-sensitive window — degradation kinetics — is easy to miss unless a person explicitly flags it.

Deliberately Omitting the Antisolvent Step

In halide-perovskite film fabrication, antisolvent quenching — dripping a nonpolar solvent during spin-coating to trigger rapid crystallization — is close to standard practice. Yet this paper explicitly notes that the films were prepared “without using an antisolvent.” The manuscript does not explicitly state the reason, but this can be attributed to the oxidation sensitivity of Sn and the phase-separation issues observed in mixed Sn-Pb perovskites.

It is an exception to the general protocol whose significance is much easier to recognize with direct experience in this materials system.

Subtle Differences in Spin-Coating and Annealing Conditions

Step	Perovskite Layer	Dopant Layer
Spin-coating	5000 rpm, 100 s	2000 rpm, 45 s
Annealing	90°C, 20 min	100°C, 10 min

The dopant layer needs to be thinner and stay confined to the surface, hence the lower rpm and shorter spin time. For annealing, the higher temperature but shorter time reflects a simultaneous set of constraints: solvent evaporation rate, the dopant’s thermal decomposition temperature, and the need to avoid recrystallizing the already-formed perovskite.

AI can faithfully reproduce “try these conditions” as text and can propose plausible physical or chemical rationales. But when those rationales are not documented, it cannot know which considerations actually led the experimenter to choose those specific values.

3. Hidden Variables Never Mentioned in the Fabrication Process

The paper simply states that all processing was done in an N2 glovebox. But a great deal hides behind that one line.

Actual O2/H2O ppm Levels in the Glovebox

SnI2 is highly susceptible to Sn2+ → Sn4+ oxidation. Actual O2/H2O levels vary from glovebox to glovebox. Even within the same glovebox, they vary day to day depending on the purge cycle and load-lock wait time. This can become a hidden source of batch-to-batch variability.

The Data-Filtering Criterion the Paper Discloses About Itself

At the very end of the Experimental Section, we state that we ultimately analyzed only the runs in which the undoped control samples showed consistently low carrier density, in the range of roughly 10¹⁴ to 10¹⁵ cm⁻³. This allowed us to rule out the possibility that the shift in carrier density arose incidentally from the sample’s exposure to oxygen.

In effect, this means that on days when glovebox conditions were poor and even the undoped sample’s carrier density spiked, that run was excluded from analysis. Where exactly to draw that threshold isn’t in any standard manual. It is a judgment built up through repeated measurements and experience with what counts as a normal versus an anomalous batch.

To get an AI to judge whether a given dataset is trustworthy, someone first has to put this batch-filtering criterion into words for it.

4. Results and Discussion: Data You Can’t Take at Face Value

The most technically interesting parts of this paper are actually the places where we openly acknowledge the limits of our own measurements.

Hall Measurement — Figure 5

The paper notes that when one configuration produced unclear signals, likely because of contact issues, only the results from the other configuration were used. In a six-terminal Hall bar measured in two configurations, if one side looks off, it gets attributed to a contact problem and discarded. Judging when a signal is “unclear,” without any quantitative threshold, is exactly the kind of hands-on know-how that’s hard to formalize. Building a robotic automation platform around this would require labeled training data for a quality classifier.

At least initially, those labels would likely need to come from experienced experimentalists who know what failure modes look like in practice.

TOF-SIMS — Figure 6, “Assessment of Doping Efficiency”

The paper states plainly that TOF-SIMS cannot generally be considered quantitative without a standard.

The dopant quantification process turns out to require several layers of correction simultaneously:

The secondary-ion yield of F⁻ can vary depending on the matrix, represented by the SnI2⁻ and PbI2⁻ signals.
The F4TCNQ samples had rougher surfaces, with aggregates visible by SEM. It therefore took longer sputtering to reach the substrate signal, making the matrix-ion intensity appear lower than in the undoped sample.
That lower matrix-ion signal must not be interpreted as a higher doping concentration. Instead, it must be attributed correctly to surface roughness.
The authors verify this directly in Figures S5 and S9 by comparing integrated intensities between films of matched thickness.
Ultimately, the Mo compound is quantified by ICP-MS, and the F4TCNQ signal is calibrated indirectly against it using the known F/Mo ratio in the Mo sample.

This correction logic comes from recognizing instrument-specific failure modes that are difficult to infer from the raw data alone and are often learned through repeated experience with the instrument.

If an AI interpreted this TOF-SIMS data automatically without accounting for matrix effects or sputter-rate differences caused by surface roughness, it could conclude that “doping concentration is X times higher” purely from the raw intensity ratio. That would be the wrong conclusion.

XPS — Figure 7

The paper states that because all measured core levels are referenced to the C 1s peak, absolute Fermi-level shifts between doped and undoped samples cannot be reliably assessed from the XPS data. The conclusions therefore have to remain limited to chemical-state information.

Since every XPS peak is calibrated against C 1s at 284.8 eV, the data cannot be used to judge absolute Fermi-level shifts before and after doping. Instead, we use only the splitting of the Sn 3d peak — 486.47 eV and 487.51 eV, a gap of about 1.0 eV — as chemical-state evidence of a Sn2+/Sn4+ oxidation-state change.

Knowing exactly where to draw the line between what a measurement supports and what it does not often comes from repeated experience with the instrument and its referencing limitations.

Film Thickness: The Baseline for Every Calculation

Both conductivity and doping-efficiency calculations depend entirely on film thickness, 270 nm, via:

ρ = Rs · t

Depending on exactly where that thickness was measured — near a scratch edge versus the center of the film — and exactly where the profilometer probe traced, the final conductivity value can shift meaningfully. This is a detail the paper never specifies. It rests entirely on the physical precision of whoever performed the measurement.

AI can readily apply the equations once the inputs are given. The harder problem is deciding how trustworthy those inputs are—a judgment that often depends on knowing how the measurement was physically made and where its uncertainty comes from.

5. Things Never Written Down at All, but That Still Change the Result

There are also variables that don’t appear anywhere in this paper, but that anyone trying to reproduce the same protocol will inevitably run into:

Glass-substrate pretreatment: plasma etching time affects surface wetting properties and, in turn, initial nucleation density during spin-coating.
Micro-gaps between the shadow mask and substrate: during Au electrode deposition, these affect electrode-edge sharpness, then effective Hall-bar channel width, then systematic error in conductivity calculations.
Epoxy curing degree: measuring before full cure allows residual solvent to affect device characteristics.
Pipetting precision: volumes such as 53 μL and 100 μL affect film-thickness uniformity. For robotic automation, even dispense speed and height need to be standardized for reproducibility.

In this study, these experimental conditions were kept consistent across repeated experiments; the point here is that changes in any of them could introduce variability when the protocol is reproduced under different conditions.

None of these variables get fully written down in any paper, largely because even the experimenter handles most of them unconsciously.

But this is often exactly the answer to:

“Why can’t I reproduce this paper exactly in my own lab?”

So, Where Does AI’s Role End?

Having walked through this paper with AI from start to finish, the boundary came through pretty clearly.

What AI Is Good At

Building first-order hypotheses from things like energy-level alignment
Accurately parsing protocol text into a reproducible form
Correctly applying equations such as doping efficiency and resistivity
Synthesizing multiple papers to generate candidate molecules or experimental conditions

Where Experimental Intuition Still Matters

Reinterpreting a failed hypothesis through second-order variables, such as actual surface behavior: solubility, aggregation, and adsorption
Knowing the why hidden behind protocol numbers such as concentration, temperature, and time, especially time-sensitive windows like precursor aging or dopant aggregation kinetics
Judging how much a given measurement can be trusted: TOF-SIMS matrix effects, Hall-contact quality, the limits of XPS referencing, and batch-filtering criteria based on glovebox conditions

All three of these are things that rarely get written explicitly into a paper.

They exist as tacit knowledge inside the experimenter’s head.

If you wanted to port this knowledge into an automated robotic platform or an AI experimental-design tool, someone would first have to do the work of making that tacit knowledge observable—through explicit rules, annotations, examples, or experimental records that an AI system can learn from.

And the person who can do that work is you: the one right now checking the O2 level in front of the glovebox, listening for a strange sound from the spin-coater, and sensing that something’s off about a TOF-SIMS signal.

One Thing to Try

Next time you read a paper or write up a protocol, try just one thing:

Write down, even in a single line: “Why is this condition this specific number?”

That’s the first step toward making tacit experimental knowledge explicit, so that AI can become a better tool for designing, running, and interpreting experiments.