What does "irreproducible" actually tell you?
The logical gap that opens up everything else.
Here is something both proponents and critics of solid-state fusion agree on: the experiments are hard to reproduce. Run an electrolysis cell one week and you might see a burst of apparent excess heat. Run it again the following week and you get nothing. For most of the scientific world, that pattern settled the question long ago. An effect you can't summon reliably probably isn't real.
The reasoning seems airtight — but it hides an assumption. It says: if the physics were real, the result would repeat; it doesn't repeat; therefore the physics isn't real. That holds only if the experiment you ran the second time was actually the same experiment as the first. And whether two runs are "the same" is not a physics question. It's a measurement question. It depends on what you controlled, what you logged, and what you only thought you held constant.
That distinction is why this article exists. Solid-state fusion (also called condensed matter nuclear science, or CMNS) lives or dies on the quality of its instrumentation — not on its theories, which remain genuinely unsettled, and not on anyone's conviction, but on whether a handful of small, slow, awkward signals can be measured well enough to mean something.1 That is a data-acquisition problem in the most literal sense. And it is an unsolved one in ways that should interest anyone who has ever wrestled with a noisy sensor.
One condition needs to be stated up front, because the whole argument rests on it. Reframing "irreproducible" as a measurement problem only works when you can actually name the uncontrolled variable — not just gesture at the possibility that some variable exists. What follows rests on a case where the variable can be named. That is a stronger claim than "just measure better," which would quietly assume every hidden variable is equally tractable. It isn't guaranteed to be.
How a 2004 government review actually came out
Not the burial most people remember.
In December 2004 the U.S. Department of Energy assembled eighteen expert reviewers to look at the evidence again, fifteen years after a first, dismissive review. The panel neither endorsed the field nor buried it. On the specific question of whether the experiments produced excess heat, the reviewers split roughly evenly.2 On whether anything nuclear was happening, about two-thirds were unconvinced — though one found the case compelling, and the rest fell somewhere between.3
The common shorthand — "the DOE was unconvinced" — misrepresents this. On excess heat, reviewers volunteered more favorable than unfavorable comments, and the panel was nearly unanimous that well-designed individual proposals deserved funding.4 The skepticism was real. So was the recommendation to keep looking.
What almost everyone on the panel did agree on was the complaint about methods: the work was poorly documented, backgrounds weren't consistently controlled, and much of the apparatus fell well short of what the question deserved.5 You could read that as an indictment of the field. Read it more carefully and it's a list of measurement failures — documentation, background control, instrument capability — none of which speak to whether the underlying physics exists. All of them speak to whether you'd know if it did.
A field can fail to reproduce a result for two entirely different reasons. The effect may not be there. Or the effect may be there, but governed by a variable nobody was measuring, so that "same setup" was an illusion both times. Telling those two cases apart is the job of instrumentation. Until you've done it, "irreproducible" describes your equipment more than it describes nature.
The variable nobody was watching
A concrete example of how a missing channel looks like a dead effect.
The most useful finding to emerge from careful work in this field came from a long electrochemical program at SRI International. It reported that excess heat in palladium-deuterium systems doesn't appear until the palladium has absorbed deuterium past a particular threshold — a specific ratio of deuterium atoms per palladium atom — and that this loading level is difficult to reach and harder to hold. Reported6 The group's conclusion was blunt: without continuously measuring that loading, an experimenter has no basis to judge whether a given run could have produced heat at all.
Caveat: This threshold comes from a single research program; independent replication of the result is limited, and whether it holds up broadly is exactly the kind of open question this whole argument turns on.
Follow the logic anyway, as a thought experiment. If the loading threshold is real, a lab that doesn't measure it will sometimes hit it by luck — and see heat — and more often fall short — and see nothing — and will record both as "the same experiment." The pattern that would result is precisely what the field is criticized for: sporadic, burst-like, maddeningly intermittent heat signals. On this reading, that intermittency isn't evidence against the effect. It's the fingerprint of an uncontrolled variable, which is to say a missing channel in the data-acquisition system.
Notice what this is and isn't. It doesn't say the heat is nuclear; it says nothing about mechanism. The claim is narrower and more durable: some fraction of the field's famous irreproducibility is plausibly a measurement artifact. The way to find out is not to argue about theory. It's to instrument the loading in real time, on every cell, and see whether heat tracks it. That experiment is mundane. It is also, in a sense, decisive — and it has rarely been done to a modern standard across many labs at once.
Three familiar problems in an unfamiliar place
The measurement challenges decompose into things instrumentation engineers already know.
Heat you can trust
The original excess-heat claims came from a particular calorimeter design — one that measured temperature at a single point in a poorly stirred cell. Critics showed, in detail, how such a setup manufactures a fake excess: the single sensor catches a local hot spot, not the true cell average, and an improperly anchored calibration does the rest. A separate analysis found that applying one fixed calibration constant to flow data, when that constant actually drifts run to run, produces an error term large enough to mimic the claimed signals. Established7
These critiques are correct — and they're useful. Each one names a specific artifact with a specific engineering fix: distributed temperature sensing, active stirring, calibration tied to a reaction of known heat output, per-run calibration tracking. Better calorimeters were built in response: flow calorimeters and Seebeck calorimeters that integrate all heat leaving an enclosure, rather than sampling one corner of it. The disagreement over excess heat is, at bottom, a disagreement over calorimeter design and calibration discipline. That kind of disagreement can be resolved.
Particles against a background
The other signature researchers look for is charged particles — energetic fragments that would indicate a nuclear reaction had occurred. The detector of choice has been CR-39, a clear plastic that records a tiny pit wherever an energetic particle passes through it. You etch the plastic to develop those pits, then count them under a microscope.
The trouble is that etching also raises surface features that aren't particle tracks, and ordinary radiation from the environment leaves tracks of its own. Historically, reading the result has meant a human at a microscope — slow work, prone to unconscious bias.
Two developments have changed the picture. Differential designs, where a detector sensitive to the claimed signal sits beside a near-identical control that isn't, let you subtract the background instead of arguing about whether it accounts for your count. And track images are now being classified by trained neural networks: one recent model sorted proton, triton, and helion tracks at 97 to 99 percent accuracy when benchmarked against expert hand-counts, and resolved overlapping tracks that a conventional threshold tool merged into one or threw away entirely.8 Track counting has become a signal-processing pipeline.
A small signal in a long record
The deepest challenge is the most generic. The interesting signals here are small and slow — a fraction of a watt of excess power riding on watts of input, or a faint track rate rising over weeks. Pulling a weak signal out of a long, drifting, noisy record is the founding problem of signal processing. The tools the situation demands are exactly the classical ones: careful averaging, lock-in detection where applicable, rigorous uncertainty propagation, and honest separation of systematic from random error.
The field's results have too often appeared without the full uncertainty budget a physicist in any other domain would insist on. Supplying that discipline isn't a favor to solid-state fusion. It's the price of any defensible measurement — and it's what would turn a contested anomaly into either a solid result or a clean refutation.
What the field can offer back
A bridge that only carries traffic one way is just a loading dock. It's worth being concrete about what instrumentation science stands to gain from engaging with this problem, beyond goodwill.
The measurement regime is awkward in genuinely useful ways. You are asked to resolve a small thermal signal against a large input, hold a calibration stable for weeks, and discriminate rare events from a structured background — all at once. That combination is a hard test bench for anything you build. A calorimeter design or a track-classification model that survives this problem will survive a lot of easier ones.
The clearest recent illustration is the Google-funded program that ran roughly four hundred experiments and found no cold-fusion effect, yet produced careful materials and measurement techniques that outlast the null result. Established9 That program deserves an honest accounting, because it tests the logic of this whole article. If good instrumentation is supposed to settle the question, then the best-instrumented campaign to date is not merely methodological progress — it is evidence, and it points toward absence. A reasonable prior should shift down in response.
The caveat that keeps it from being decisive is the same loading-threshold point from above: the team worked with nickel-hydrogen systems and a range of palladium-deuterium samples, and reported it could not confirm it had reached the extreme loading regimes that proponents say are necessary. A fully instrumented null is strong evidence. It isn't proof.
In 2023, ARPA-E put roughly ten million dollars into eight projects aimed at testing whether these reactions are real.10 Several are explicitly diagnostics work — one team's stated goal is to build the capability to measure hypothetical neutron, gamma, and ion emissions; another brings accelerator-grade detector expertise to quantify event rates.11 Funding signals institutional interest, not a verdict on the physics. The agency framed the goal as breaking a long stalemate — and a stalemate broken by better measurement is a data-acquisition story whichever way the physics falls.
Two questions, not one — and one reason for real skepticism
What instrumentation can settle, and what it can't.
Before closing, a distinction the title quietly blurs needs to be made explicit. "Is there an anomalous heat signal" and "is that signal from nuclear fusion" are different questions. The first is a calorimetric and statistical question — the kind instrumentation can genuinely close. The second is a question about mechanism that no calorimeter answers by itself.
There is a physics objection that better measurement does not touch. Known nuclear physics predicts that if deuterium nuclei were fusing at the rates implied by the claimed heat levels, the reaction would produce neutrons and tritium in quantities large enough to be immediately apparent — and dangerous. Reports of helium-4 without the expected neutrons or gamma rays would require the reaction to follow branching ratios that differ from everything else physicists have ever measured by orders of magnitude.12 No data-acquisition system resolves that. It is a real reason for skepticism, and it sits outside the scope of what better instrumentation can settle.
So the honest position is this: a set of anomalous signals has been reported for over thirty-five years; the objections most resolvable by instrumentation are objections about measurement quality; and the measurement has rarely been done to the standard the question deserves. Contested The narrow claim — that anomalous heat might be real — is instrument-resolvable. The broader claim — that the heat is nuclear fusion — requires more than calorimetry.
The prescription is narrow and comes with a stated condition for being wrong. Apply to these cells the same rigor you'd bring to any hard measurement: pre-registered protocols, traceable calibration, blinding where bias can creep in, open data, multi-lab replication on shared hardware. Instrument the loading variable continuously. Subtract backgrounds rather than debating them. Report the full uncertainty budget.
A fully instrumented null — loading measured throughout, backgrounds subtracted, uncertainty accounted for — counts as evidence that the effect isn't there. A sustained series of such nulls would settle it in the negative. The measurement frame has to cut both ways or it isn't honest.
Run that gauntlet and one of two things happens. The anomalies dissolve into well-understood artifacts, and a long-standing question finally closes. Or they survive contact with first-rate instrumentation, and the people who built that instrumentation find themselves at the center of something genuinely strange. Both outcomes run through the data-acquisition system.
Notes
- On condensed matter nuclear science as an instrumentation-limited field, see U.S. Department of Energy, Report of the Review of Low Energy Nuclear Reactions (Washington, DC: DOE, December 2004), summarized in "US review rekindles cold fusion debate," Nature News, December 2, 2004, nature.com. ↩
- "US review rekindles cold fusion debate," Nature News, December 2, 2004: the panel was reported "split approximately evenly" on whether the experiments produced excess heat. ↩
- "Back to Square One," Scientific American, reviewing the December 2004 DOE report: on nuclear origin, roughly two-thirds of reviewers found the evidence unconvincing, one found it compelling, and the remainder were somewhat convinced. scientificamerican.com. ↩
- On the excess-heat question, contemporaneous accounts report reviewers volunteered more favorable than unfavorable comments, and that the panel was nearly unanimous in recommending that well-designed individual proposals deserve funding. See P. L. Hagelstein, "Constraints on energetic particles in the Fleischmann–Pons experiment," and the field history in "Gatekeeping: a Partial History of Cold Fusion," arXiv:2601.09996. ↩
- Reviewers also cited deficiencies in data interpretation and noted equipment well short of the state of the art; see Scientific American, "Back to Square One," and the Nature News account, both cited above. ↩
- M. C. H. McKubre and F. L. Tanzella, "Cold Fusion, LENR, CMNS, FPE: One Perspective on the State of the Science," Journal of Condensed Matter Nuclear Science 4 (2011): 32–44. The SRI program reported excess-power onset above a deuterium-to-palladium loading threshold, and concluded that without measuring that loading, an experimenter has no basis to judge whether a run could have produced heat. jcmns.org. ↩
- On calorimeter artifacts: "An Assessment of Claims of 'Excess Heat' in 'Cold Fusion' Calorimetry," Thermochimica Acta 315 (1998), sciencedirect.com; and "A systematic error in mass flow calorimetry demonstrated," Thermochimica Acta (2002), sciencedirect.com, which showed that a global calibration constant applied to run-varying data produces an error term approaching the magnitude of reported signals. ↩
- Y. Wang, A. X. Chen, et al., "Deuterium–deuterium fusion charged particle detection using CR-39 and a deep learning model," Radiation Measurements (2025), sciencedirect.com. A YOLOv8 network trained on noise-reduced images classified protons, tritons, and helions at 97–99% accuracy against expert counts on the reported test images, and resolved overlapping tracks that a threshold tool merged or discarded. ↩
- C. P. Berlinguette et al., "Revisiting the cold case of cold fusion," Nature 570 (2019): 45–51. The program reported no excess-heat effect across its campaign while contributing measurement and materials methods; the team noted it could not confirm it had reached the highest loading regimes some proponents emphasize. ↩
- U.S. Department of Energy / ARPA-E, "U.S. Department of Energy Announces $10 Million in Funding to Projects Studying Low-Energy Nuclear Reactions," February 17, 2023, arpa-e.energy.gov. ↩
- University of Michigan project described in the ARPA-E listing as providing "capability to measure hypothetical neutron, gamma, and ion emissions from LENR experiments"; and "Berkeley Lab to Lead ARPA-E Low Energy Nuclear Reactions Project," atap.lbl.gov. ↩
- On the canonical physics objection: known deuteron–deuteron branching ratios predict that fusion at the claimed heat levels would yield readily detectable neutron and tritium fluxes, and reports of helium-4 without commensurate neutrons or gamma rays would require branching ratios departing from observation by orders of magnitude. See the summary in P. L. Hagelstein, "Models for nuclear fusion in the solid state," arXiv:2501.08338. ↩
Editorial note: This article presents a scholarly synthesis of SSF's relationship to signal processing and data acquisition. The underlying nuclear claims of SSF/LENR remain scientifically contested. Evidence claims are tiered as established, contested, or reported-but-unconfirmed as noted inline. Readers are directed to primary experimental literature for empirical evaluation.
