best-of-atomistic-machine-learning: a ranked index of 520 atomistic ML projects

A generated list, not a curated opinion

The Best of Atomistic Machine Learning list is described as a ranked list of awesome atomistic machine learning projects, updated regularly, and it contains 520 open source projects with a total of 250K stars grouped into 23 categories.

The mechanism behind the ranking is stated in the readme. All projects are ranked by a project-quality score, which is calculated based on various metrics automatically collected from GitHub and different package managers. The badge in the header links to the best-of-generator documentation for the exact definition of that score, and the list itself points at a whole family of best-of lists with a link to the tooling that creates them.

That distinction matters more here than it would in a list of web frameworks. In a field where the number of serious tools is small and their differences are subtle, a popularity ranking can mislead: the most-starred interatomic potential code is not automatically the one with the best accuracy, and a project that is scientifically important but not maintained on GitHub can be invisible to a metric built on GitHub activity. The list is a map of the ecosystem, and it is a good one, but the word ranked here means ordered by measurable activity, not ordered by merit.

The list declares its own bias honestly: the current focus is more on simulation data rather than experimental data, and weighted towards physical systems rather than drug design, though contributions from other fields are warmly welcome. The homepage link points at go.fzj.de/baml, and there is a Zenodo DOI for citation.

The 23 categories, from interatomic potentials to language models

The category breakdown is the most useful thing in the list, because it maps the field rather than the popularity contest.

Interatomic Potentials is the largest section with 78 projects, followed by Representation Learning with 63 and Datasets with 55. Then come Density functional theory (ML-DFT) with 39 and Community resources with 39, Educational Resources with 33, Representation Engineering with 31, Language Models with 25 and General Tools with 24. The smaller sections are Universal Potentials with 19, Generative Models with 18, Molecular Dynamics with 15, Materials Discovery and Mathematical tools with 14 each, and Unsupervised Learning with 10.

The tail is revealing: Active learning has 9 projects, Wavefunction methods (ML-WFT) 7, Electronic structure methods (ML-ESM) 6, Visualization 6, Data Structures 4, Explainable Artificial intelligence 4, Reinforcement Learning 2, and Probabilistic ML and Others 1 each.

Read as a map of the field, that distribution says something specific. The mature, crowded area is potentials and representations, and the frontier areas with one or two projects are wavefunction methods, explainability and reinforcement learning. The single-entry categories are also a reminder that the list reflects what has an open source implementation, not what matters scientifically: probability weighting and uncertainty quantification in atomistic ML have a substantial literature and very few tools here.

For a newcomer, the Educational Resources and Community resources sections are the way in. For someone with a specific need, the Data Structures section being only four projects is itself a signal about where the field concentrates.

The marker legend, and what the icons actually tell you

Each entry carries a set of markers, and the legend is where the list's method is most transparent.

A medal indicates the combined project-quality score, a star is the star count from GitHub, and a trend arrow shows whether a project is trending up or down. A plus sign marks a project that was recently added. The people, fork, issue and last-update markers give the contributors count from GitHub, the fork count from GitHub, the issue count from GitHub and the last update timestamp on the package manager. A downloads marker gives the download count from the package manager, and a package marker gives the number of dependent projects.

Three markers encode age and activity directly, and they are the ones worth reading first. A chick marks a new project, less than 6 months old. A sleeping symbol marks an inactive project, 6 months with no activity. And a skull marks a dead project, 12 months with no activity.

Those three are the list's most useful signal, because they tell you the age of the tool and whether anyone is still fixing it. A skull next to a project you were about to adopt is a warning that no one has touched it in a year, and in a research field where the underlying methods and the supporting libraries move quickly, a year of no activity often means incompatibility with current versions of everything else.

The dependent projects count is the other one to look at. A library with many dependents is a load-bearing part of somebody's workflow, which is a different and often more meaningful kind of popularity than a star count.

What an entry looks like, and where to get the code

Entries are structured so that the install path is one copy away, because most projects in this field are distributed several ways and a reader usually wants the specific one.

The top entry in the Active learning section is DP-GEN, the deep potential generator, with a combined quality score of 25 and 400 stars, under LGPL-3.0, tagged ML-IAP, MD and workflows. Its entry lists the GitHub repository with 76 contributors, 200 forks, 2.1K downloads, 8 dependent projects and 370 issues at 3 percent open, and it gives three install routes. The GitHub route is a clone:

git clone https://github.com/deepmodeling/dpgen

The PyPI route is a pip install:

pip install dpgen

And the Conda route is:

conda install -c deepmodeling dpgen

The same three-way pattern recurs. FLARE, the second entry, is an open source Python package for creating fast and accurate interatomic potentials, MIT licensed, with 44 contributors, 78 forks and 10 downloads, and it also lists a conda channel.

That the same project appears with a pip, a conda and a source install is not a duplication but a useful signal about how the community in this field distributes code. A package that is on both pip and conda is a project that expects to be used inside an existing scientific environment, and the download counts on those package managers, which the entries record with their own update timestamps, are a truer measure of usage than the GitHub numbers, since a research code is often pip-installed in a cluster environment rather than cloned.

Contributing, and the history directory behind the list

The contribution path is unusually open for a curated list. You can open an issue, submit a pull request, or directly edit projects.yaml, which is the data file the whole thing is generated from.

That last option is the one to notice. The presence of a projects.yaml at the top level, alongside a config/ directory, means the list is not hand-maintained Markdown. It is generated output, and projects.yaml is the source of truth. The upstream best-of tooling provides a create-best-of-list guide for anyone who wants to make their own list this way.

The consequences are mostly good. A generated list can be regenerated on a schedule, which is why the readme says updated regularly, and the releases are dated snapshots: 2025.04.08, 2025.04.09 and 2025.04.09-13.10 are the most recent ones listed. A history/ directory, a latest-changes.md and a CITATION.cff file are also present, so the list can be cited and its evolution inspected.

The licence is CC-BY-SA-4.0, which is the right choice for a document like this: attribution plus share-alike, so someone can fork the list for a subdomain of the field and has to keep it open in the same way.

The trade-off of generated lists is that the editorial voice is a formula. Every entry gets the same markers and the same treatment, so the list cannot tell you that a project is brilliant but unmaintained, or that two projects are alternatives for the same task. It tells you the shape of the ecosystem, and that shape is a genuinely useful thing to have in a field with hundreds of tools.

What the ranking does not measure

The limitation is the one inherent to the method, and it is worth being precise about it because the medal emoji invites over-reading.

The score is computed from metrics collected automatically from GitHub and package managers. That means stars, forks, contributors, issue counts, download counts, dependent project counts and update timestamps. Every one of those measures activity and adoption. None of them measures whether a potential is accurate, whether a workflow is correct, whether a dataset has the right provenance, or whether a paper's claims replicate.

For a field with fast-moving methods and a large volume of unreviewed preprints, that gap is significant. Two projects can have nearly identical scores while one implements a method with a known systematic error and the other is careful about it. The list cannot tell you that, and the marker legend does not pretend to.

The list is also not a mirror of the whole field. Its own statement of focus, more on simulation data than experimental data and more on materials than drug design, means the 520 projects skew accordingly, and the single-project categories show how thin coverage is in some areas. The topical tags reinforce this, spanning ai4science, computational-chemistry, condensed-matter, density-functional-theory, drug-discovery, electronic-structure, interatomic-potentials, molecular-dynamics, quantum-chemistry, scientific-computing, scientific-machine-learning and surrogate-models.

Read the ranking as a way to find candidates, then read the actual repositories. That is exactly how the list is meant to be used, and the 520 entries with their install commands are an unusually good starting index.

Editorial conclusion

This list fits a researcher who already knows the field and needs a map of which tools are popular, recently updated and still alive, rather than a beginner who needs to be told which method to use. Before trusting a ranking, read the entry for the project in question and check the last update timestamp and the dependency count yourself, because the score is built from repository activity rather than from the quality of the science, and a dead-project marker is a statement about commit history.

Frequently asked questions

Official sources

1. JuDFTteam/best-of-atomistic-machine-learning on GitHub
2. License: CC-BY-SA-4.0
3. Project website
4. README
5. Releases