Making Rowan Easier To Use
collaboration in organizations; predicting slow jobs; myriad ui/ux improvements; faster conformer searching; new AIMNet2; vDZP preprint
We’ve just released a whole suite of changes focused on making Rowan easier to use and more powerful. Briefly: we’ve launched new organization-specific features, lots of website improvements, better documentation, a new conformer-search workflow, and a better version of AIMNet2 (and we also published a preprint). If you’re curious about any of this, read on!
New Organization-Specific Features
We’re excited to be launching our first features to support organization-level administration and sharing needs.
In Rowan, organizations are set up with role-based permissions. Rowan organization admins can set credit usage warnings and overage limits at the account level as well as track, monitor, and control activity within their organization. This gives organizations control over their billing, real-time observability, and access to each user’s historical data in perpetuity.
Organizational users can also access a shared team drive in the Rowan platform, allowing users to collaborate on projects. Moving a workflow into the shared drive grants organization-wide access, letting people within an organization view results securely, leave notes on each other’s work, and collaborate on computational projects—all without having to deal with the hassle of import/export, filetype compatibility, and software versioning.
Traditionally, computational chemistry has been largely a “single-player” activity: sharing results within a computational team is tough, and communicating results to scientists outside the team is even harder. We’re trying to change this paradigm and make it possible to bring computational insights to teams, not just to people. If you’re a part of a fast-moving research organization and want to partner with us to explore what this might look like for you, reach out!
Various Website Improvements
In response to user feedback, we’ve been making lots of changes to our website.
One of our big priorities is helping scientists understand how long a given calculation is expected to take. This is in general a very challenging question—but we’ve built a set of internal heuristics that perform pretty well at binning jobs into “fast” or “slow.” Rowan now displays an alert if a job is predicted to take longer than 5 minutes to run:
In the future, we’re hoping to be able to more accurately forecast how long a job will take.
Other changes include:
Optimizations and scans now have a “play” button to automatically show all steps in sequence.
Users can now file bug reports or feedback from within the platform by clicking the “bug” button in Rowan’s top navigation bar.
The new “bug” button for reporting issues and leaving feedback Searching in Rowan has gotten much more powerful—you can now filter by workflow type, by status, and for starred items.
Rowan’s advanced search capabilities The molecule editor now has a “Help” menu, accessible via an icon at the top of the molecule editor. Plus, you can copy-and-paste molecules now.
The molecule editor “Help” menu We’ve added a “freeze atoms” constraint type that allows the Cartesian position of various atoms to be frozen during optimizations.
Markdown formatting is now supported in the “Notes” field of folders and calculations.
Improved Documentation
In our last update, we emphasized the importance of documentation:
We’ve been conducting a lot of user meetings, and something that we’ve heard a lot recently is that our documentation needs to be way better. While six months ago Rowan was very simple, Rowan is now becoming much more capable and complex, and it’s easy to get confused by a growing number of poorly documented features—particularly for scientists without previous experience in computational chemistry. (emphasis added)
We now have 8 new video tutorials, in addition to the 5 we shared last time:
Redox potential prediction (video, transcript; 6 minute runtime)
Torsional scans (video, transcript; 7 minute runtime)
Descriptors and PCA (video, transcript; 4 minute runtime)
pKa prediction (video, transcript; 9 minute runtime)
Multistage optimization (video, transcript; 5 minute runtime)
Spin-state prediction (video, transcript, 3 minute runtime)
Bond-dissociation energies (video, transcript; 2 minute runtime)
Tautomer search (video, transcript; 7 minute runtime)
We’ve also created new scientific overviews for each workflow, similar to what might be found in the SI of a scientific paper. You can read these overviews in the “Science” section of our documentation site.
As always, our documentation is a work in progress—we have a lot more improvements in the pipeline, and welcome any feedback from our users!
A Faster Conformer Workflow
Our previous conformer-search workflow worked well overall, but suffered from a few consistent issues: in particular, duplicate conformers were frequently missed, leading to identical conformations in the output and inefficient computations. We’ve totally revamped our conformer search code, and are now using the efficient + intelligent conformer screening utilities in CREST for all conformer searches, which should lead to noticeable improvements in performance and accuracy. (For more details on exactly what we’re doing, you can check out our detailed scientific explanation.)
A Better Version of AIMNet2
A new version of AIMNet2, one of Rowan’s workhorse NNPs, was released earlier this year. The new AIMNet2 release has slightly improved accuracy on a variety of benchmarks and, crucially, now works on periodic systems. We’ve updated the version of AIMNet2 that Rowan uses, so all AIMNet2-based calculations will be slightly more accurate, and AIMNet2 can now be used for periodic systems.
Preprint: Exploring the vDZP Basis Set
As Rowan users know, we’re big fans of Stefan Grimme’s “composite”/”3c” DFT methods, which use specially optimized functionals, basis sets, and empirical corrections to achieve high accuracy while running significantly faster than traditional DFT methods. Unfortunately, it’s difficult to develop arbitrary composite DFT methods, since each one is bespoke and requires a concerted scientific effort and extensive parameterization.
Our recent preprint (arXiv) shows that the new vDZP basis set can essentially turn arbitrary DFT methods into “composite” methods by minimizing basis-set errors at the double-ζ level. We compare vDZP-based methods to other double-ζ basis sets and existing composite methods like B97-3c and r2SCAN-3c, and examine combining vDZP with other functionals to create new DFT methods that are fast and accurate.
We’re very excited about this work, and hope to incorporate more vDZP-based methods into Rowan in the future.