Historical research guide · 0804 · 2024
Exploring artificial intelligence techniques to research low energy nuclear reactions
Anasse Bari · Tanya Pushkin Garg · Yvonne Wu · Sneha Singh · David Nagel
In one minute
Anasse Bari and Tanya Pushkin Garg's group at New York University, working with David Nagel of George Washington University, turn modern language AI on a problem that is not about language at all — the scattered literature of Low Energy Nuclear Reactions. LENR is the field of nuclear reactions run at ordinary temperatures inside metal lattices, and thirty-five years of it sit in conference papers, newsletters and preprints nobody has read as one body of work. The team assembled 3,424 documents, more than 200,000 pages, and let unsupervised topic models sort them without being told what to look for. The best combination — BERTopic with Microsoft's e5-base-v2 sentence embeddings and KMeans clustering — resolved the corpus into six recurring themes, every one of them about deuterium, palladium lattices, excess heat or the particles that come out. They also built and deployed LENRsim, a public tool that finds the papers most like one you hand it, at roughly a quarter of the brute-force runtime. The work was funded by the Anthropocene Institute.
Connection
Chapter 12 is the lattice-confinement fusion chapter, and this paper is the map of the territory it sits in — the first time the whole LENR corpus has been read as one dataset rather than a pile of separate results. It is also chapter 1's evidence ladder applied mechanically: six themes emerge from the literature itself, and every one of them names a material, a measurement and a laboratory you can go and check.
From the source
The LENR literature from the late 1980s to the present can be assembled into a single searchable corpus of 3,424 unique documents spanning more than 200,000 pages, built from the 4,743-entry LENR-CANR bibliography plus 1,250 papers and reports retrieved by hand from Google Scholar, ScienceDirect, Nature, IEEE Xplore and preprint repositories.Section 3, Dataset
Of the topic models tested, BERTopic paired with Microsoft's e5-base-v2 sentence embeddings and KMeans clustering gives the best joint result on the LENR corpus — topic coherence 0.14 and topic diversity 0.98 — outperforming both Latent Dirichlet Allocation, which scores 0.06 coherence, and Top2Vec with Doc2Vec, which reaches 0.99 diversity but negative coherence.Section 5.3.2, Table 2
Unsupervised clustering of the corpus, with the search terms LENR and cold fusion deliberately removed as stop words so they cannot steer the result, resolves the whole field into six recurring themes: neutron emission and deuterium transmutation; deuterons and electrons interacting in lattice structures; the hydrogen-palladium electrochemical system; the Fleischmann excess-heat experiment; electrolysis and tritium production at a palladium cathode; and calorimetry of excess heat in electrolytic cells.Section 5.3.2, Table 3; Section 4
A two-phase retrieval algorithm — match the query first against each cluster's representative document, then only against the members of the winning cluster — reaches an average cosine similarity of 0.8925 against the brute-force algorithm's 0.9, at an average runtime of 11.86 milliseconds against 46.03, roughly one quarter, and the advantage grows with corpus size.Section 6.2, Table 4
LENRsim, the resulting similarity tool, is built in React and Flask, deployed publicly at lenrdashboard.com, and accepts a PDF or plain text query; the next version is to index full paper texts rather than titles and abstracts alone, which is the limit imposed here by BERT token lengths.Section 6.3, Dashboard; Section 7, Conclusion
The authors' case for the field: LENR harnesses nuclear reactions at ambient temperatures, offering cost-efficient energy production, minimal radiation emission during operation, no significant radioactive waste and negligible greenhouse gas emissions — which is why they argue data-driven tools should now accelerate both its research and its commercialisation.Section 1, Introduction
The way in
https://doi.org/10.3389/frai.2024.1401782How to cite it
Anasse Bari, Tanya Pushkin Garg, Yvonne Wu, Sneha Singh, David Nagel (2024) Exploring artificial intelligence techniques to research low energy nuclear reactions doi:10.3389/frai.2024.1401782