Pavlo O. Dral

I am looking forward to participating in Faraday Discussion's Molecular excited states theory and experiment, 14-16 September 2026, Cambridge, UK, rsc.li/excitedstate...

Deadline for Oral abstract submissions is 15 December 2025.

#compchem #aichem #mlchem

A Descriptor Is All You Need: Accurate Machine Learning of Nonadiabatic Coupling Vectors Nonadiabatic couplings (NACs) play a crucial role in modeling photochemical and photophysical processes with methods such as the widely used fewest-switches surface hopping (FSSH). There is, therefore, a strong incentive to machine learn NACs for accelerating simulations. However, this is challenging due to NACs’ vectorial, double-valued character and the singularity near a conical intersection seam. For the first time, we design NAC-specific descriptors based on our domain expertise and show that they allow learning NACs with never-before-reported accuracy of R2 exceeding 0.99. The key to success is also our new ML phase-correction procedure. We demonstrate the efficiency and robustness of our approach on a prototypical example of fully ML-driven FSSH simulations of fulvene targeting the SA-2-CASSCF(6,6) electronic structure level. This ML-FSSH dynamics leads to an accurate description of S1 decay while reducing error bars by allowing the execution of a large ensemble of trajectories. Our approach is generalizable to more states as we demonstrate for a three-state ML-FSSH simulation of methylenimmonium cation. Our implementations are available in open-source MLatom.

Just out in JPCL — our accurate ML approach for nonadiabatic coupling vectors!

This work took years — from early ML-FSSH struggles to finding physics-based descriptors (energy gradient differences) and improving MLIPs
#compchem @mbarbatti.bsky.social

doi.org/10.1021/acs....

Delighted to present our AI-driven #compchem work at ICCOC 2025, Shenzhen. Huge congrats to my PhD student Xinxin for winning the Best Poster Prize on AIQM methods — she is surely one of the brightest up-and-coming scientists!

same for our AIQM methods which are based on xTB

Nice work by my co-supervised PhD student Mateusz!

You can read the work at doi.org/10.1021/acs.... .

Continuing the previous post, here is one of my favorite examples of how things can go wrong when you use universal #ML potentials - MD of H2. I love to show this example to my students, and it is in my online course (aitomistic.com/en/sub/course) at @aitomistic.com .
#compchem

YouTube video by Prof. Pavlo O. Dral SMLQC 2025, Pavlo O. Dral's talk "Universal AI Models for Ground and Excited States"

My talk at @Smlqc1Smlqc -2025 is now online.
This is the third SMLQC edition ( www.smlqc2025.com )!
Talk is covering #ML models for #compchem simulations, also available with #AIagents at the @aitomistic.com Hub ( aitomistic.xyz ).
youtu.be/gIpE_pqF2e4

Poster on Aitomia presented by Hassan Nawaz at #MDMM25, where @pavlodral.bsky.social also gave a talk on Aitomia.

Showcasing Aitomia's ability to autonomously design #compchem workflows with #AIagents, such as calculating reaction thermochemistry and spectra, on aitomistic.xyz

2/2 If you are interested in the course on @aitomistic.com , it is available online for free for academic users. You can also pre-register for a living & course with webinars. See more at aitomistic.com/en/sub/course

#compchem #mlchem #aichem

Are neural scaling laws leading quantum chemistry astray? Neural scaling laws are driving the machine learning community toward training ever-larger foundation models across domains, assuring high accuracy and transferable representations for extrapolative t...

1/2Just came across this preprint discussing #ML potentials' failure even for H2.

In my course, I have been showing this to my students already for many years, with both astonishing examples of failures of popular foundational ML models and tutorials on how to solve them.

arxiv.org/abs/2509.26397

Are neural scaling laws leading quantum chemistry astray? Neural scaling laws are driving the machine learning community toward training ever-larger foundation models across domains, assuring high accuracy and transferable representations for extrapolative t...

1/2Just came across this preprint discussing #ML potentials' failure even for H2.

In my course, I have been showing this to my students already for many years, with both astonishing examples of failures of popular foundational ML models and tutorials on how to solve them.

arxiv.org/abs/2509.26397

Thank you for sharing!
You can run #compchem simulations with AIQM2 as described in our tutorials: mlatom.com/docs/tutoria...
Also, online via a web browser on @aitomistic.com Hub at aitomistic.xyz (free)

hard work by Xinxin (the first author), she has many more such models in her library!
You can run #compchem simulations with AIQM2 as described in our tutorials: mlatom.com/docs/tutoria...
Also, online via a web browser on @aitomistic.com Hub at aitomistic.xyz (free)

All-in-one leaning is a very handy method to learning from multiple levels of theory (and data sources in general) simultaneously. Better than alternative transfer learning in many respects. Just out in JCTC: pubs.acs.org/doi/10.1021/...

🚀 MLatom 3.18.3 released!
Updates include:
✨ MACE-OFF model zoo interface
✨ Raman spectra plots (Python)
✨ Extended ML model for 2-photon absorption spectra

📄 Notes: mlatom.com/docs/release...

🔗 aitomistic.xyz

It is always nice to see creative ways the users apply our methods and software (UAIQM & #MLatom) to solve their #compchem problems:

www.sciencedirect.com/science/arti...

You can use them online too at the @aitomistic.com
Hub.

Back in 2021, I wrote about a future where computers could autonomously run & analyze #compchem simulations: shorturl.at/Vq4tq
Now, I’m thrilled to be building #AIagents that make this vision real!

YouTube video by Prof. Pavlo O. Dral MACE-OFF: OFF or ON POINT? ML force field from JACS available on Aitomistic Hub

MACE-OFF is a popular universal #ML potential for #compchem recently published in JACS. Is it up to the hype? In the video on YouTube, we dissect its strengths and weaknesses. You can try MACE-OFF yourself on @aitomistic.com Hub.

youtu.be/4Hv-Sus1384

thank you, Ellis, that would be awesome!

It is one of the most frequently used methods on the
@aitomistic.com Hub for online simulations via a web browser, including the AI assistant Aitomia.

AIQM2: organic reaction simulations beyond DFT Density functional theory (DFT) is the workhorse of reaction simulations, but it often suffers from either prohibitive cost or insufficient accuracy. In this work, we report AIQM2—the universal AI-enh...

Link to the paper: doi.org/10.1039/D5SC...
Also see blog post: dr-dral.com/chemical-sci...

AIQM2 just got published in @chemicalscience.rsc.org !

This #ML method's high speed, competitive accuracy, and robustness enable organic reaction #compchem simuls beyond what is possible with the popular DFT methods. It can be used for TS opt and dynamics, often with chem. accuracy.

#AI + atomistic, #compchem, simulations evolve so fast I have to redo my hands-on materials multiple times a year 🤯
That's a continuously updated Living Course is the way to go.

Only a few days left for the early-bird registration to the Symposium on Machine Learning and Quantum Chemistry #SMLQC 2025 www.smlqc2025.com !

Organized by the one and only Konstantinos Vogiatzis

#ml #compchem #mlchem #aichem

Theoretical study on accurate and affordable molecular IR spectra calculations with the AIQM methods available on our Aitomistic Hub (aitomistic.xyz) was recently published in J. Phys. Chem. A.
Paper: doi.org/10.1021/acs....
Video recap: youtu.be/hkzM5qC8njI
#compchem #mlchem #aichem

Talk ✅ #iupac2025 #compchem #mlchem #aichem

Just landed in Kuala Lumpur to attend #IUPAC2025. If you are there and going to the conference dinner or otherwise want to meet - drop me a message ☺️

Did you know that you can run advanced #ML and common DFT #compchem calculations on Aitomistic Hub at www.aitomistic.xyz via simple MLatom input files?
Submitting MLatom Python scripts or launching Jupyter notebook also works!

Detailed tutorials at mlatom.com/docs .

Two of my PhD students graduated!
Lina Zhang did an incredible job of progressing the ML surface hopping dynamics.
Fuchun Ge developed novel NN methods directly predicting MD trajectories, and did lots of work on MLatom and ML potentials.
Each of them have 10+ publications!
dr-dral.com/people