Solid State Fusion & Machine Learning + AI

What Happened When Google Tried to Settle This

A $10 million null result that was more interesting than a null result usually is.

In 2015 a team with funding from Google set out to put the old cold-fusion controversy to rest. They spent about ten million dollars over several years, drew researchers from MIT, UBC, Maryland, and Lawrence Berkeley, and ran careful experiments. In 2019 they published their findings in Nature. They found no excess heat.¹

So far: not surprising. The claim that certain metals loaded with hydrogen isotopes release unexplained heat has been contested since it burst into headlines in 1989, and most physicists would have predicted that outcome. What made the Google team's paper unusual was the sentence that followed.

They reported that their main difficulty was not evaluating the physics. It was getting the material into the right state to test the physics at all. The claim requires packing deuterium (a heavy isotope of hydrogen) into a palladium metal lattice at very high concentrations, close to one deuterium atom for every palladium atom. Reaching and confirming that concentration turned out to be harder than expected. The standard method for measuring how much had gone in — watching how the metal's electrical resistance changed — proved unreliable. They eventually switched to a structural technique: shining X-rays through the sample and watching the lattice expand as it filled up.²

Their conclusion was not "the effect doesn't exist." It was closer to: we are not sure that anyone, including ourselves, has reliably explored the corner of parameter space where the claimed effect is supposed to live.³

Read that as a physicist and it sounds like a hedge. Read it as someone who studies optimization and inference and it sounds like a precise description of an underpowered search.

Why This Is Your Problem as Much as Theirs

Strip away the nuclear question. What's left has familiar names.

The field these experiments belong to goes by several names: cold fusion, low-energy nuclear reactions, condensed matter nuclear science. Whether any of the claimed effects are real is contested, and this article does not take a position on that question. What is worth examining is the shape of the experimental problem, because the shape is something a CS student will recognize.

An experimenter chooses a material, a loading method, a temperature schedule, an electrochemical protocol, and a dozen other parameters. They run the experiment for days or weeks and look for a small, intermittent signal against a drifting background. Most runs show nothing. The ones that do report something rarely reproduce cleanly.⁴

Decompose that into subproblems and you get: a large, expensive-to-sample parameter space that needs to be searched efficiently; anomaly detection in a noisy time series with non-stationary baseline; surrogate modeling of an experiment too costly to run exhaustively; and calibrated uncertainty quantification on a claimed effect near the noise floor. These are not physics problems dressed up in ML clothing. They are ML problems, and the question of what produces the heat (if anything does) does not change any of them.

If the effect is real, these are the tools that would find and characterize it. If it is not, these are the tools that would show that clearly. Either answer would be worth having.

Four Places Where the Methods Connect to the Physics

Each of these has published mainstream precedents. None has been applied to this problem yet.

Note: Every method below requires data to run on. Usable, standardized data from this field barely exists yet. Read this section as what becomes possible once the data infrastructure exists, not as four things you could run on a laptop this semester.

Searching the Loading Space with Bayesian Optimization

The Google team's central frustration — reaching and confirming a specific loading regime — is a sequential experimental design problem. When each experiment costs weeks and the parameter space is large, you cannot grid-search it. The standard alternative is to build a surrogate model, typically a Gaussian process, and let an acquisition function decide which point in the space to evaluate next, trading off exploration against exploitation.

This is Bayesian optimization, and it has become established as the engine of what researchers call "self-driving" materials laboratories. One such system, called Ada, used a closed-loop Bayesian optimizer to tune the processing of thin-film solar-cell materials, adjusting composition and annealing conditions across many interacting parameters to find configurations that a human experimenter would have been slow to reach.⁵,⁷ The chemistry is entirely different from palladium-deuterium, but the structure of the problem is the same: an expensive experiment, many coupled knobs, and an optimizer deciding what to try next.

Nobody has pointed this machinery at the loading question in solid-state fusion. That is an open opportunity, not a published result.

Treating Excess Heat as a Detection Problem

An excess-heat claim is, operationally, a claim that a calorimeter's output departs from what its inputs predict, by a small amount, for a while, sometimes. Deciding whether that departure is real or artifactual is signal detection against a non-stationary background: the instrument drifts, mundane chemical processes can mimic the signal, and the events of interest are sparse and unlabeled.

Machine learning has mature tools for this regime. Online anomaly detection methods can adapt as the background distribution shifts over time.⁶ Closer to the physics, deep time-series anomaly detection has been built and tested on calorimeter data from high-energy physics experiments, where the challenge of separating rare real events from instrument artifacts is structurally similar.⁸

One honest constraint belongs here: anomaly detection is data-hungry. What blocks this application is not the existence of suitable methods but the availability of clean, labeled, standardized calorimetry records from this field. That is an argument for building the data infrastructure, not against the approach.

Simulating Hydrogen in the Lattice

The behavior solid-state fusion is concerned with — hydrogen isotopes packed into a metal lattice at extreme concentration — is hard to simulate from first principles. Quantum-mechanical methods are accurate but scale poorly; they become prohibitively slow for the system sizes and timescales that matter.

Machine-learned interatomic potentials are designed to close that gap. The idea is to train a neural network on quantum-mechanical calculations for a range of atomic configurations, then use the trained network to run molecular dynamics at something close to quantum accuracy but at a fraction of the computational cost. This is now established practice in computational materials science.⁹

The connection to this problem is close to direct. A 2022 study built exactly such a neural-network potential for the palladium-hydrogen system and used it to compute how protium, deuterium, and tritium diffuse through palladium at temperatures ranging from cryogenic to above the melting point, including the quantum effects that cause the three isotopes to behave differently.¹⁰ The work was published as mainstream computational materials physics with no connection to cold fusion. That framing is precisely why it serves as a credible bridge: the lattice dynamics it describes are a natural starting point for any mechanistic discussion of what solid-state fusion claims happens inside the metal.

One premise worth making explicit: this approach assumes that lattice-level dynamics are the right description of whatever is (or is not) occurring. That is a reasonable modeling choice, not a proven one. A genuine effect could in principle involve physics that a classical interatomic potential, however well trained, does not capture.

Putting Principled Error Bars on the Answer

A field whose central evidence is a small, intermittent signal cannot afford loose error estimates. Whatever a calorimeter or particle detector reports, the question that matters is: how confident should we be that this is not an artifact?

Ensemble methods and Bayesian uncertainty quantification give principled ways to attach calibrated confidence intervals to a model's predictions. Both are now established in scientific machine learning, and have been applied to exactly the relevant type of problem: predicting how hydrogen interacts with metal surfaces, where the reaction probabilities are small and the cost of a wrong answer is high.¹¹

In a field this contested, the deliverable is not a louder claim. It is a number with an honest interval around it.

What This Field Gives Back to ML

An adversarial test bed that most benchmark datasets cannot supply.

Solid-state fusion data occupies a corner that machine learning rarely gets to work in. Small signals. Long runs. Severe non-stationarity. Sparse, disputed labels. And a strong prior, held by most of the scientific community, that any positive result is an artifact until proven otherwise.

Most published anomaly-detection benchmarks are nothing like this. They suffer from unrealistically high anomaly density and labels that are not trustworthy,¹³ which means a method that scores well on them has not necessarily been tested hard. A regime where a false positive can end a career and a false negative might mean missing a genuine energy source is a demanding environment for methods that usually get graded on tidier data.

Worth thinking about carefully: this adversarial prior inverts a bias researchers usually worry about. In most fields the concern is that publication pressure inflates positive results. Here the selection pressure runs the other way: in a stigmatized area where reporting a marginal positive is professionally risky, the incentive is to suppress or ignore such results rather than publish them. Positive reports that do survive that filter are not obviously products of optimistic bias. That is an unusual feature of the environment, and one that makes reasoning about selection effects somewhat cleaner than in a typical field.

There is also a cautionary lesson the field hands back. When an autonomous lab at Berkeley reported synthesizing dozens of new materials in seventeen days using ML-guided active learning, the published abstract described 41 of 58 targets as successfully produced.¹² Readers then challenged whether the materials were truly new, and a correction followed: "novel" was reframed as meaning new to the platform rather than new to the literature, one compound was removed because it had been in the training data, and the success count was revised.

None of this is an argument against autonomous discovery. The rigor solid-state fusion actually needs — pre-registered targets, conservative novelty claims, independent replication — is the same rigor that ML-driven discovery sometimes skips. A field that has spent thirty years being dismissed for overclaiming is, in an odd way, useful training for a discipline learning how easy it is to overclaim with a model in the loop.

What Would Actually Settle This

The offer to a skeptical computational scientist.

The most useful contribution machine learning can make here is not a theory of the heat. What it can supply is the apparatus of a search done properly: shared datasets with enough metadata to make runs from different labs comparable; surrogate-guided experimental design so that scarce, expensive cells are spent where they are most informative; drift-aware detection with stated false-positive rates; calibrated uncertainty on every number reported; and independent, ideally blinded, replication coordinated across sites.⁴ None of that requires believing the effect is real. It requires only that the question is considered worth answering, and that it be answered with instruments sharp enough to trust the result.

The original claim was reportedly dismissed within about forty days of its announcement in 1989.¹⁴ Forty days is not how long it takes to search a high-dimensional space carefully. The tools to do that search now exist. Running it well would resolve the question in one direction or the other, and both directions are publishable. The only outcome that is not is a continued refusal to look carefully.

What remains open: whether usable, standardized data can be assembled; whether the loading regime the claims require can be reached reliably enough to test against; and whether the relevant physics, if there is any, operates at the level a machine-learned potential can describe. Those are the threads worth pulling on.

Editorial note: This article presents a scholarly synthesis of the relationship between solid-state fusion and machine learning methods. The underlying nuclear claims of SSF/LENR remain scientifically contested. The ML methods discussed are established; their application to SSF is proposed, not yet published, unless noted otherwise. Readers are directed to primary literature for empirical evaluation.

Notes

1. Curtis P. Berlinguette et al., "Revisiting the cold case of cold fusion," Nature 570 (2019): 45–51. A perspective synthesizing a multi-institution program spanning UBC, MIT, Maryland, LBNL, and Google; approximately US$10 million over several years; no excess heat detected within the parameter space tested. ↩
2. Berlinguette et al., Nature 570 (2019). The program reached loading near D/Pd ≈ 0.96, still short of the ≥ 0.875 regime the original claims pointed to, and adopted X-ray measurement of lattice expansion after resistance-based estimates proved unreliable. ↩
3. Berlinguette et al., Nature 570 (2019), 45: the authors conclude that continued skepticism is warranted but that the relevant conditions require further investigation before the phenomenon can be ruled out. ↩
4. U.S. Department of Energy, Report of the Review of Low Energy Nuclear Reactions (Washington, DC: DOE, 2004). Reviewers were roughly evenly split on the excess-heat evidence; the report identified intermittency and reproducibility as central obstacles and called for standardized multi-lab protocols. ↩
5. A. Gilad Kusne et al., "On-the-fly closed-loop materials discovery via Bayesian active learning," Nature Communications 11 (2020): 5966, https://doi.org/10.1038/s41467-020-19597-w. Reports approximately a tenfold reduction in experiments required to map a materials system; establishes Gaussian-process surrogates with active-learning acquisition as standard practice. ↩
6. Mengmeng Zhang et al., "Online anomaly detection with sparse Gaussian processes," Neurocomputing 403 (2020): 383–399, https://doi.org/10.1016/j.neucom.2020.04.077. Cited as established ML capability; application to SSF calorimetry is proposed, not yet published. ↩
7. B. P. MacLeod et al., "Self-driving laboratory for accelerated discovery of thin-film materials," Science Advances 6, no. 20 (2020): eaaz8867, https://doi.org/10.1126/sciadv.aaz8867. Ada optimized hole mobility of an organic hole-transport material for perovskite solar cells via a Bayesian optimizer. Cited as a methodological parallel for closed-loop optimization; the materials system is not palladium and the example is not an SSF result. ↩
8. "Synthetic Data Generation with Lorenzetti for Time Series Anomaly Detection in High-Energy Physics Calorimeters," arXiv:2509.07451 (2025). Demonstrates deep time-series anomaly detection on physics-calorimeter data; the SSF-calorimetry application is proposed, not published. ↩
9. Pascal Friederich et al., "Machine-learned potentials for next-generation matter simulations," Nature Materials 20 (2021): 750–761, https://doi.org/10.1038/s41563-020-0777-6. Survey of the field. ↩
10. Hajime Kimizuka, Bo Thomsen, and Motoyuki Shiga, "Artificial neural network-based path integral simulations of hydrogen isotope diffusion in palladium," Journal of Physics: Energy 4 (2022): 034004, https://doi.org/10.1088/2515-7655/ac7e6b. ML potential for Pd–H; computes protium/deuterium/tritium diffusion across 50–1500 K including nuclear quantum effects. Mainstream materials physics, no LENR framing. ↩
11. W. G. Stark et al., "Machine Learning Interatomic Potentials for Reactive Hydrogen Dynamics at Metal Surfaces Based on Iterative Refinement of Reaction Probabilities," Journal of Physical Chemistry C (2023), https://doi.org/10.1021/acs.jpcc.3c06648. Ensemble learning with full uncertainty quantification for hydrogen-on-metal reaction probabilities. ↩
12. N. J. Szymanski et al., "An autonomous laboratory for the accelerated synthesis of inorganic materials," Nature 624 (2023): 86–91, https://doi.org/10.1038/s41586-023-06734-w. A 2026 author correction (Nature 650 (2026): E1) reframed "novel" as new to the platform rather than new to science, confirmed 36 of 40 reported successes reached the correct outcome, and removed one compound that had been in the training data. ↩
13. Discussion of unrealistic anomaly density and unreliable labels in common time-series benchmarks: OML-AD study, arXiv:2409.09742 (2024), citing Wu and Keogh (2021). ↩
14. Same as note 4 above; cited separately for the data-infrastructure argument. ↩
15. Michael McKubre, as quoted in Discover (2023): by his account, the initial dismissal of the 1989 claim came within roughly forty days. Offered as one participant's recollection, not a neutral history; the broader response, including the first DOE review, continued for months and years afterward. ↩

Introduction: The Search Problem Nobody Has Run This Way

Here is a fact that should bother a machine-learning researcher more than it bothers most physicists. In 2015 a team funded by Google set out to settle, once and for all, whether the old cold-fusion claim had anything to it. They spent about ten million dollars over several years, pulled in people from MIT, the University of British Columbia, the University of Maryland, and Lawrence Berkeley National Laboratory, and ran careful experiments. In 2019 they published a Nature Perspective summarizing the program. They found no excess heat.¹

And then they said something stranger than a null result usually warrants. They reported that the hardest part was not the physics. It was getting the material into the state the original claim required at all. Packing deuterium into palladium past about seven atoms for every eight of palladium turned out to be extremely difficult, and measuring how much had actually gone in was its own problem. The standard trick, watching electrical resistance change, gave them errors, so they switched to shining X-rays through the metal to watch it swell.² Their conclusion was not "the effect is not real." It was closer to "we are not sure anyone, including us, has reliably reached the corner of parameter space where the claim says to look."³

Read that as a physicist and it sounds like an excuse. Read it as someone who works on optimization and inference and it sounds like a confession: the field has been doing an underpowered search of a high-dimensional space with noisy, drifting instruments and almost no shared data. That is a description of a problem your discipline was built to solve.

Section I — The Bridge Is the Search, Not the Physics

You do not have to take a position on cold fusion to find the methodology interesting.

Solid-state fusion, also called condensed matter nuclear science, is the claim that certain metals loaded with hydrogen isotopes occasionally release more heat than any chemical reaction can account for, and sometimes show nuclear signatures. Whether that claim is true is contested and not the subject of this piece. What is not contested is the shape of the problem. An experimenter chooses a material, a loading method, a temperature schedule, an electrochemical or gas-phase protocol, and a dozen other knobs, runs the cell for days or weeks, and watches a small, intermittent, hard-to-reproduce signal against a drifting background. Most runs show nothing. The ones that report something rarely repeat cleanly.⁴

A fair objection lands first. Thirty-odd years and well over three thousand papers is not obviously an underpowered search. The careful version of the claim is narrower and survives the objection: nobody has run a well-instrumented, ML-guided, shared-data search of this space. The Nature authors' own word was "underexplored," not "unexplored," and that distinction is the one worth keeping.

Strip the nuclear question away and what remains is a stack of problems with familiar names. A search over a large, expensive-to-sample parameter space. Anomaly detection in noisy time series with baseline drift. Surrogate modeling of an experiment too costly to run exhaustively. Uncertainty quantification on a claimed effect near the noise floor. These are not metaphors for machine-learning problems. They are machine-learning problems, and the heat source's origin does not change any of them. If the effect is real, these tools are how you would find and characterize it. If it is not, these are the tools that would show that cleanly and put the question to rest. Either outcome is worth a graduate student's time.

Section II — Where Your Methods Touch the Metal

Four contact points, each tied to published mainstream work. One caveat first: every method below consumes data, and the usable, standardized SSF data it would consume barely exists yet. Read this section as what becomes possible once the data infrastructure of Section IV is built, not as four things you could run tomorrow.

Searching the loading space

The Google team's central frustration — reaching and confirming a specific loading regime — is a sequential experimental-design problem. When each experiment is expensive and the space is large, you do not grid-search it. You build a surrogate, usually a Gaussian process, and let an acquisition function pick the next experiment to maximize information gain. This is Bayesian optimization, and it is now the standard engine of autonomous "self-driving" materials labs, where it has cut the number of experiments needed to map a materials system by roughly an order of magnitude.⁵ ESTABLISHED. One such platform, Ada, used a Bayesian optimizer in a closed loop to tune the processing of thin-film materials for solar cells, modifying composition and annealing conditions to find configurations a human search would have been slow to reach.⁷ The chemistry there is not solid-state fusion's chemistry, but the machinery is exactly transferable: a costly experiment, many interacting knobs, and an optimizer deciding what to run next. Nobody has yet pointed it at the loading-versus-excess-heat question. [PROPOSED — the application to SSF loading is an open opportunity, not a published result]

Excess heat as a detection problem

An excess-heat claim is, operationally, the claim that a calorimeter's output departs from what its inputs predict, by a small margin, for a while, sometimes. Deciding whether that departure is real is signal detection against a non-stationary background, because calorimeters drift, recombination and other mundane processes mimic the signal, and the interesting events are sparse and unlabeled. Machine learning has mature tools for exactly this regime: online anomaly detection that adapts to concept drift,⁶ and, closer to the physics, deep time-series anomaly detection has been built and tested on high-energy-physics calorimeter data with deliberately injected defects.⁸ ESTABLISHED. One honest constraint belongs here rather than buried in a footnote: anomaly detection is data-hungry, and what blocks this application is not the availability of methods but the availability of clean, labeled, standardized SSF calorimetry to train and validate on. That is an argument for Section IV, not against the proposal. [PROPOSED — mainstream methods exist; their use on SSF calorimetry is not yet in the literature, and is gated by data]

Simulating hydrogen in the lattice

The behavior that solid-state fusion cares about — hydrogen isotopes packed into a metal lattice at extreme concentration — is hard to simulate. First-principles quantum methods are accurate but too slow for the system sizes and timescales that matter. This is the gap that machine-learned interatomic potentials were invented to close: train a neural network on quantum-mechanical data, then run molecular dynamics at near-quantum accuracy and a fraction of the cost.⁹ ESTABLISHED. A 2022 study built a neural-network potential for palladium-hydrogen alloys and used it to compute how protium, deuterium, and tritium diffuse through palladium across a wide temperature range, including the nuclear quantum effects that distinguish the three isotopes.¹⁰ That work was done as mainstream materials physics, with no cold-fusion framing at all, which is precisely why it is a credible bridge. The lattice dynamics it models are a natural starting point for any mechanistic discussion of solid-state fusion. That last clause carries a premise worth owning: it assumes lattice dynamics is the right level of description for whatever, if anything, is happening. It is a reasonable modeling choice, not a proven one, and a real effect could in principle live in a regime such a potential does not capture.

Putting a number on the uncertainty

A field whose central evidence is a small, intermittent signal cannot afford casual error bars. Whatever a calorimeter or detector reports, the load-bearing question is how confident we should be that it is not an artifact. Ensemble methods and uncertainty quantification — now standard equipment in scientific machine learning — give principled ways to attach calibrated confidence to a model's output, and have been used to do exactly that for reaction-probability predictions of hydrogen interacting with metal surfaces.¹¹ ESTABLISHED. In a field this contested, the deliverable is not a louder claim. It is a defensible probability with an honest interval around it.

Section III — What Solid-State Fusion Gives Back

A bridge that only carries traffic one way is a courtesy, not a collaboration.

The return is not exotic, and that is the point. Solid-state-fusion data sits in a corner that machine learning rarely gets to work in: small signals, long runs, severe non-stationarity, sparse and disputed labels, and a strong adversarial prior that any positive is an artifact until proven otherwise. Most published anomaly-detection benchmarks are nothing like this; they suffer from unrealistic anomaly density and unreliable labels, a complaint the time-series community makes about its own datasets.¹³ A regime where the cost of a false positive is a career and the cost of a false negative might be a missed energy source is a genuinely demanding test bed for methods that usually get graded on cleaner data.

That adversarial prior is worth dwelling on, because it inverts an assumption machine-learning researchers usually bring to messy data. The normal worry is publication bias inflating positive results. Here the selection pressure runs the other way. In a stigmatized field where a false positive can end a career, the incentive is to suppress marginal positives, not to manufacture them. So the positive reports that do survive are not obviously the product of optimistic file-drawer bias; if anything, the filter has been working against them. That is an unusually clean setting for reasoning about selection effects, and it is a point that cuts in the field's favor more than its defenders tend to press.

There is also a cautionary lesson the field hands back, and it is worth stating plainly because it cuts against easy enthusiasm. When an autonomous lab at Berkeley reported synthesizing dozens of new compounds in seventeen days using machine learning and active learning, the headline figure — 41 of 58 targets — came straight from the peer-reviewed abstract. That is what makes it instructive. Readers then disputed whether the compounds were genuinely new to science, and a formal correction followed: it reframed "novel" as new to the platform rather than new to the literature, confirmed the system had reached the right outcome in 36 of 40 reported successes, and removed one compound that had slipped in from the training data.¹² CONTESTED. The episode is not an argument against autonomous discovery. It is an argument for the exact rigor solid-state fusion needs and machine learning sometimes skips: pre-registered targets, conservative claims about novelty, independent checking. And the overstatement lived in the paper itself, not just a press release, which is the more sobering version of the lesson. A field that has spent thirty years being dismissed for overclaiming is, perversely, good training for a discipline learning how easy it is to overclaim with a model in the loop.

Section IV — The Reproducibility Frame Is the Honest One

If there is a single offer to make a skeptical computational scientist, it is this one.

The most useful thing machine learning can give solid-state fusion is not a theory of the heat. It is the apparatus of a search done right. Shared, standardized datasets with the metadata to make runs comparable. Surrogate-guided experimental design so that scarce, expensive cells are spent where they are most informative. Drift-aware detection with stated false-positive rates. Calibrated uncertainty on every reported number. Independent, ideally blinded, replication coordinated across labs.⁴ None of that requires anyone to believe the effect is nuclear, or even that it is real. It requires only that the question is considered worth settling rather than worth avoiding, and that it be settled with instruments sharp enough to trust the answer. Notice too that this section is upstream of Section II: the data infrastructure described here is what makes the four methods there usable in the first place.

That framing should appeal to the part of a researcher that dislikes an unfinished problem more than it fears an unfashionable one. The original claim was, by one participant's account, reported and largely dismissed within forty days.¹⁴ Forty days is not how long it takes to search a high-dimensional space carefully. We now have the tools to do the search the way it should have been done the first time. Running it well would resolve the question in one direction or the other. Both answers are publishable. Only the refusal to look is not.

Editorial note: This article presents a scholarly bridge between machine learning methodology and solid-state fusion as a search problem. The underlying nuclear claims of SSF/LENR remain scientifically contested. Evidence and method claims are tiered as established, contested, or proposed as noted inline. Readers are directed to the primary literature for empirical evaluation. Draft status: DRAFT–UNREVIEWED pending technical sign-off.

References & Footnotes

1. Curtis P. Berlinguette et al., "Revisiting the cold case of cold fusion," Nature 570 (2019): 45–51. A Nature Perspective synthesizing a multi-institution program (UBC, MIT, Maryland, LBNL, Google); roughly US$10M over several years; no excess heat detected within the parameter space tested. ↩
2. Berlinguette et al., Nature 570 (2019); the program reported reaching loading near D/Pd ≈ 0.96, still short of the ≥ 0.875 regime the original claims pointed to, and adopted X-ray measurement of lattice swelling after resistance-based loading estimates proved unreliable. ↩
3. Berlinguette et al., Nature 570 (2019), 45: the authors conclude that continued skepticism is warranted but that the relevant conditions require further investigation before the phenomenon can be ruled out. ↩
4. On intermittency, reproducibility difficulty, and the case for standardized multi-lab protocols: U.S. Department of Energy, Report of the Review of Low Energy Nuclear Reactions (Washington, DC: DOE, 2004); reviewers were roughly evenly split on the excess-heat evidence. ↩
5. A. Gilad Kusne et al., "On-the-fly closed-loop materials discovery via Bayesian active learning," Nature Communications 11 (2020): 5966, https://doi.org/10.1038/s41467-020-19597-w. Reports a roughly tenfold reduction in the experiments required to map a materials system; establishes Gaussian-process surrogates with active-learning acquisition as standard practice. ↩
6. Mengmeng Zhang et al., "Online anomaly detection with sparse Gaussian processes," Neurocomputing 403 (2020): 383–399, https://doi.org/10.1016/j.neucom.2020.04.077. Cited as established ML capability; application to SSF calorimetry is proposed, not yet published. ↩
7. B. P. MacLeod et al., "Self-driving laboratory for accelerated discovery of thin-film materials," Science Advances 6, no. 20 (2020): eaaz8867, https://doi.org/10.1126/sciadv.aaz8867. Ada optimized the hole mobility of spiro-OMeTAD via a Bayesian optimizer. Cited as a methodological parallel for closed-loop optimization of a multi-knob, expensive experiment; the materials system is not palladium and the example is not an SSF result. ↩
8. "Synthetic Data Generation with Lorenzetti for Time Series Anomaly Detection in High-Energy Physics Calorimeters," arXiv:2509.07451 (2025). Demonstrates the method class on physics-calorimeter time series; the SSF-calorimetry application is proposed. ↩
9. On machine-learned interatomic potentials generally: Pascal Friederich et al., "Machine-learned potentials for next-generation matter simulations," Nature Materials 20 (2021): 750–761, https://doi.org/10.1038/s41563-020-0777-6. ↩
10. Hajime Kimizuka, Bo Thomsen, and Motoyuki Shiga, "Artificial neural network-based path integral simulations of hydrogen isotope diffusion in palladium," Journal of Physics: Energy 4 (2022): 034004, https://doi.org/10.1088/2515-7655/ac7e6b. ML potential for Pd–H; computes protium/deuterium/tritium diffusion across 50–1500 K including nuclear quantum effects. Mainstream materials physics, no LENR framing. ↩
11. W. G. Stark et al., "Machine Learning Interatomic Potentials for Reactive Hydrogen Dynamics at Metal Surfaces Based on Iterative Refinement of Reaction Probabilities," Journal of Physical Chemistry C (2023), https://doi.org/10.1021/acs.jpcc.3c06648. Ensemble learning with full uncertainty quantification for hydrogen-on-metal reaction probabilities. ↩
12. N. J. Szymanski et al., "An autonomous laboratory for the accelerated synthesis of inorganic materials," Nature 624 (2023): 86–91, https://doi.org/10.1038/s41586-023-06734-w. The peer-reviewed abstract states the A-Lab "realized 41 novel compounds from a set of 58 targets" in 17 days. A 2026 author correction (Nature 650 (2026): E1) reframed "novel" as new to the platform rather than new to science, confirmed the platform reached the correct outcome in 36 of 40 reported successes, and removed one compound mistakenly present in the training data. Presented as a cautionary, CONTESTED example; the overstatement originated in the peer-reviewed paper, not only in press coverage. ↩
13. On the poor quality of common time-series anomaly benchmarks (unrealistic anomaly density, unreliable labels): discussion in the OML-AD study, arXiv:2409.09742 (2024), citing Wu and Keogh (2021). ↩
14. Michael McKubre, as quoted in Discover (2023): by his account, the initial wave of dismissal of the 1989 claim came within roughly forty days. Offered as one participant's recollection, not a neutral history; the broader back-and-forth continued for months and years afterward. ↩