AIQM2 is out: better and faster than B3LYP for reaction simulations!

AIQM2 is the long-awaited successor of the highly successful AIQM1 (see Nat. Commun. paper). It has overall improved accuracy and much better performance for transition states, where it produces high-quality geometries and barrier heights: it is better than B3LYP/6-31G* but orders of magnitude faster! And AIQM2 is surely much better than its baseline GFN2-xTB while retaining its speed. See our preprint in ChemRxiv for theoretical background and benchmarks.

Take for example this pericyclic reaction: barrier height at AIQM2 is much closer to the gold-standard CCSD(T)/CBS (error less than 2 kcal/mol) while B3LYP/6-31G* and AIQM1 are way off and GFN2-xTB completely failed even find the correct transition state (the purple value is for another TS found by GFN2-xTB instead):

This high quality of the AIQM2 allowed us to run overnight (no kidding!) a thousand quasi-classical trajectories downhill from the transition state and revise the B3LYP-D3 product distribution reported in JACS in 2017:

An important advantage of AIQM2 is that it can also be simply installed via pip while AIQM1 has known frustrating problems with obtaining and installing the required third-party packages. Just update your MLatom installation via

pip install--upgrade mlatom

or simply run the AIQM2 calculations on the https://XACScloud.com.

You can use AIQM2 in the same way as B3LYP or AIQM1 – just provide the method name in the input file or your Python script as described in the extensive tutorials https://xacs.xmu.edu.cn/docs/mlatom/index.html, e.g., for geometry optimization:

Now, if you attentively followed our updates, you might ask: what about the UAIQM platform? AIQM2 is indeed just one of the models in the UAIQM library and it was the model auto-selected by the UAIQM platform for the above pericyclic reaction! We do recommend using the UAIQM platform which is much more than just a library hosting dozens of new methods such as AIQM1 and AIQM2, each of which has a different design, strengths, and weaknesses. The problem is that the scale of capabilities of UAIQM so far has gone unappreciated by traditional publishing and the UAIQM preprint unfortunately still remains only a preprint. Thus, we decided to go the traditional way too, and publish separate accounts on the composition of the individual models and know-how alongside the overarching UAIQM framework putting it all together.

To summarize, AIQM2 can be installed easily via pip and is available on GitHub but is still limited to, e.g., CHNO elements. UAIQM is more powerful, supports all periodic table, improves over time and is also available for online computations but is not available on GitHub or pip. If you want to use UAIQM locally – please contact us at contact@mlatom.com.

P.S. As you have guessed, AIQM3, AIQM4, etc. are upcoming too, many of them with elements beyond CHNO.