Chemical Science: “AIQM2: Organic Reaction Simulations Beyond DFT”

AIQM2 got published in Chemical Science! This ML method’s high speed, competitive accuracy, and robustness enable organic reaction simulations beyond what is possible with the popular DFT methods. It can be used for TS structure search and reactive dynamics, often with chemical accuracy. It is one of the most frequently used methods on the Aitomistic Hub for online simulations via a web browser, including the AI assistant Aitomia.

The accuracy is competitive with DFT and, in some cases, can even approach the gold-standard coupled cluster accuracy. The speed of AIQM2 is orders of magnitude faster, such that propagating 1k trajectories for a representative pericyclic reaction can be done overnight, when it takes 160 CPU-years with DFT. On top of that, AIQM2 provides error bars for its predictions, which provides a useful metric for researchers to judge how trustworthy the simulations are – a feature, which is absent with DFT approaches, where it is not clear how accurate the simulations are until their results are compared to a more accurate reference.

AIQM2 is based on a delta-learning approach, where a set of neural networks corrects the predictions of the popular semi-empirical quantum mechanical method GFN2-xTB (its modified version to be precise). This allows us to elevate the accuracy of GFN2-xTB to nearly coupled cluster accuracy in many cases, while retaining the robustness of this semi-empirical approach. Simulations with coupled cluster accuracy are often unaffordable due to the high cost. Robustness is a big issue with pure universal neural network potentials, which are becoming nowadays more and more popular, but still suffer from breakdowns when they are used for applications far from their training set. The combination of the fast semi-empirical method and neural networks enables predictions which are much faster than typical DFT approaches, while our benchmarks show that the accuracy is comparable and sometimes even better.

Don’t take our word for it: as one of the Reviewers of our paper said:

I used and tested a few popular MLPs and found their first-generation AIQM1 more general and reliable than the other MLPs. In addition, although a few general-purpose MLPs were developed by different groups, almost all (except very recent AIMNET2 and AIQM2) cannot apply to reaction processes and TS structures. Their MLP-predicted TS properties are very impressive.

The reviewer mentions AIQM1, which is the first generation of the successful AIQM approaches, which, however, had sub-par accuracy for transition states. This prompted us to develop AIQM2, addressing the issue.

This unique combination of speed, accuracy, and robustness for reaction simulations, enabled us performing large scale downhill quasi-classical trajectory simulations to revise the post-transition state product distribution in a representative reaction, previously explored in JACS by experts in the field using DFT B3LYP-D3/6-31G*.

We found that AIQM2 has much smaller errors in barrier heights and reaction energies compared to this DFT level: the errors of AIQM2 wrt coupled cluster reference are within 2 kcal/mol, while DFT has errors of ca 10 kcal/mol for the minima. GFN2-xTB performs even worse than DFT with the huge errors of ca 25 kcal/mol and even fails to find one of the transition states. This shows that the popular schemes where lower-level methods are used for initial reaction or conformer explorations might miss many key structures. AIQM2 removes the need for such initial explorations as it can be directly used to provide high-quality estimates.

In addition, high speed of AIQM2 enabled us to improve the precision of the simulations by propagating more trajectories (1k vs ca 100 in the original JACS study), as well as use the smaller time step (0.1 vs 1.0 fs), which is essential for total energy conservation:

If you are wondering about more extensive benchmarks, they are provided in the paper and below is a comparison for the reaction barrier heights:

AIQM2 performs well for neutral, closed-shell organic molecules with CHNO elements. It can also be applied to charged species and radicals. Extensions to more elements are available as part of the UAIQM platform (see also an overview of the AIQM models), where AIQM2 was a model UAIQM_GFN2xTB*@cc versioned 20240106. Yes, the model existed since 6 January 2024, and just recently got published! As you can see from that overview, many more models already exist, which are awaiting their publication.

AIQM2 is already used in the optimization of a system with 700+ atoms and to predict other properties, such as accurate IR spectra at high speed.

AIQM2 is available in open-source MLatom, which comes with detailed, extensive tutorials.

Paper:

• Yuxinxin Chen, Pavlo O. Dral*. AIQM2: Organic Reaction Simulations Beyond DFT. Chem. Sci. 2025, accepted manuscript. DOI: 10.1039/D5SC02802G. Open access, under the CC-BY license. The figures and text are partially adopted from this article.
  Preprint on ChemRxiv: https://doi.org/10.26434/chemrxiv-2024-j8pxp (2024.10.08).