The lab
roomtsc is a project to find a room-temperature superconductor. It uses AI to build better models of superconductivity, does original research with them, and publishes all of it: papers, model, code and data. The project is run by a lab in San Francisco that is part of ERP.AI.
The target is a superconductor at 300 K and one atmosphere. That needs three things at once. Electrons have to pair above 300 K, the pairs have to stay phase coherent above 300 K, and the material has to exist at one atmosphere and stay that way. The highest reproduced transition temperature at one atmosphere is 133 to 138 K, less than half of the target.
There are no examples to learn from, so the lab starts from what limits the transition temperature. For pairing by phonons there is a known upper scale, set by one quantity called the Hopfield sum, and the lab's model estimates that quantity from a crystal structure.
The work so far is theory and computation. The first paper gives the argument and a census of 4,619 hydrides calculated at one atmosphere. The largest hydrogen coupling among them is below what 300 K needs, and it changes severalfold with a setting of the calculation. The model, also called roomtsc, is released with its weights, code and test results. On structure types it has not seen it cuts the error of a fitted constant by about half, and it fails the test of predicting above its training range. The companion paper reports both.
The project also keeps a reference on the subject and the record of what experiments have shown. Each record entry and news item links to its source and says what the result shows and what it leaves open.
Papers
Room temperature, one atmosphere
What limits superconductivity at room temperature and one atmosphere, and how far published calculations and experiments stand from each limit.
PDF (41 pages, 2.1 MB) · Read online
roomtsc: an estimator of the Hopfield sum, tested on held-out structure types
An estimator of the Hopfield sum and of the scattering strength per proton of hydrides from crystal structure, its tests on held-out structure types, and their results.
PDF (19 pages, 0.7 MB) · Read online
Model and data
The model page runs the estimator on a structure you paste and gives its test results. The data page lists the files and scripts behind the papers.
Lab updates
- First use of the model. It was run on the 30,822 hydride compounds of the release that have no label, most of them unstable in a harmonic calculation, which the first paper names as its largest gap. It places them higher than the labelled ones on average and none at the level 300 K requires. The largest estimates are platinum hexahydrides with small instabilities, which are the first candidates for a direct calculation. The model page gives the list and its limits.
- Draft 6 of the first paper. Each of the five alkali hexahydrides at the top of the census has a stable insulating neighbour with one more alkali atom: the PBEsol database places K2PtH6, Rb2PtH6, Cs2PtH6, Rb2NiH6 and Cs2NiH6 on the convex hull with gaps of 3.1 to 3.8 eV. The hull distances of the five themselves are still unknown. The paper now counts structure types independently of the origin and says that 67 of its 4,619 records are second records of a compound.
- The companion paper on the model is online as a first draft: roomtsc: an estimator of the Hopfield sum, tested on held-out structure types (PDF). It reports both runs of the tests, the leak in the first splits, and the failed extrapolation requirement.
- The model's tests have been run again on the corrected splits. With structure types that share nothing with training, each estimator's error is 0.40 to 0.54 of the best baseline's, against a mark of 0.50: the graph network meets it with the largest values held out, both meet it on the held-out family, and both miss it narrowly over the grouped folds. The extrapolation requirement is failed again. The model page gives the table.
- The first test splits of the model leaked: a structure type could carry two labels, so some compounds of a held-out type stayed in training, and five held-out compounds had a second record in training. A second deposit labels structure types independently of the origin and keeps one record per compound. The tests are being run again on it, and both sets of results will be shown.
- roomtsc 0.1 is public: an estimator of the Hopfield sum of a crystal structure and of the scattering strength per proton of a hydride, with a page that runs it on a structure you paste and an API behind it. Three tests were run on splits deposited before training. Each of the two estimators meets the mark on two of them and misses it narrowly on the third. Neither predicts a value above its training range, so the extrapolation requirement is failed, and every result says so.
- Draft 5 of the paper replaces the census of published tables with one over the full Alexandria electron-phonon release, 4,619 hydrides at one atmosphere. The typical scattering strength per proton is 7.2 eV Å, a third of the figure in draft 4, which came from compounds already selected for a high transition temperature. The largest hydrogen Hopfield parameter is 5.6 eV/Ų, against the 8.7 to 12 that 300 K needs with a single mode, and it depends strongly on the smearing used. The model and its tests move to a companion paper. The data files and scripts are public.
- Before training the model, the lab deposited the splits its tests will use, the reference value of every held-out compound and the predictions of the baselines: manifest, splits, baselines. The labels come from 27,151 compounds of the Alexandria electron-phonon release, 4,619 of them hydrides with a separable hydrogen part.
- Draft 4 was revised after a citation audit of every section and a recheck of 1,540 printed values. About 160 corrections were made, most of them to sources and wording, and the reference list grew from 174 to 217.
- Draft 4 of the paper is online: Room temperature, one atmosphere (PDF).