Good news: someone finally built a universal neural network that can simulate how molecules behave when light hits them. Bad news: your quantum chemistry professor's job security just took a hit.

Okay, not really. But OMNI-P2x - short for "all-in-One Machine-learning Neural-network Interatomic Potential 2 for eXcited states" (scientists really need a naming workshop) - is a genuine first. Published in Nature Communications by Martyka, Tong, Jankowska, and Dral, this model does something no universal ML potential has done before: it handles excited electronic states, not just the boring ground-state stuff every other potential has been stuck on.

Wait, Why Should I Care About Excited States?

Every time a molecule absorbs a photon of light, it enters an excited state. That's the entire basis of solar cells, OLEDs in your fancy TV, photodynamic cancer therapy, and those light-activated molecular switches that make chemists lose their minds at conferences. Understanding what happens after a molecule grabs a photon is the whole ballgame for designing better versions of all those technologies.

The problem? Simulating excited states with proper quantum mechanics (specifically time-dependent density functional theory, or TD-DFT) is brutally expensive. We're talking hours of supercomputer time for a single molecule's UV/Vis spectrum. Want to screen a million candidate molecules for your next-gen solar cell? Hope you budgeted for a decade of compute.

Enter the Overachieving Pattern Matcher

OMNI-P2x was trained on 3.1 million molecules from the PubChemQC dataset, learning excited-state properties at the TD-B3LYP level. For ground states, it learned from the ANI-1ccx dataset at gold-standard CCSD(T) accuracy. The clever bit is what the authors call "all-in-one learning" - the network simultaneously trains on different levels of quantum chemical theory, essentially learning to speak multiple dialects of physics at once.

The architecture builds on the ANI neural network framework but bolts on two extra input features: a state ordering number (telling the model which electronic state you're asking about) and a one-hot encoded feature for the QM theory level. Simple additions, massive payoff.

The result? UV/Vis absorption spectra in milliseconds instead of hours. Real-time photodynamics simulations that used to require days of CPU time. And it approaches TD-DFT accuracy while running circles around traditional semiempirical methods - both faster and more accurate than the old-school shortcuts.

The Azobenzene Flex

To show off, the team used OMNI-P2x to rationally design visible-light-absorbing azobenzene systems. Azobenzenes are molecular photoswitches - they literally change shape when you shine light on them - and they're hot candidates for everything from smart materials to light-controlled drugs. Screening azobenzene derivatives for the right absorption properties used to require painstaking quantum chemical calculations for each candidate. Now you can rip through them like a deck of cards.

If you're into visualizing how these molecular architectures connect and branch, tools like mapb2.io can help you map out the relationships between different photoswitch families and their design principles - think of it as a mind map for molecular design spaces.

The Catch (There's Always a Catch)

OMNI-P2x currently supports seven elements: hydrogen, carbon, nitrogen, oxygen, fluorine, sulfur, and chlorine. That covers most of organic chemistry and a solid chunk of pharmaceutical space, but if you're working with metal complexes or heavy elements, you'll need to wait for the sequel. Fine-tuning on custom datasets is supported though, so the model is designed to be a starting point rather than a finished product.

It's also worth noting that "approaches TD-DFT accuracy" means it's not at TD-DFT accuracy. For high-stakes predictions, you'd still want to validate with full quantum chemical calculations. Think of OMNI-P2x as the metal detector - it tells you where to dig, but you still bring a shovel.

The Bigger Picture

Universal ML potentials for ground states (like MACE-MP-0 and ANI-2x) have already transformed computational chemistry. OMNI-P2x extends that revolution to photochemistry, a domain that's been waiting for exactly this kind of speed boost. The model is open-source under MIT license on GitHub and accessible through the MLatom platform, which means anyone can start using it today.

For a field where a single excited-state trajectory simulation could eat up a research group's monthly compute budget, that's not just convenient. It's the difference between studying one molecule and studying thousands. And somewhere in those thousands might be the molecular photoswitch, photocatalyst, or light-harvesting material that actually changes something.

So basically? They taught a neural network to understand what happens when molecules get excited. And honestly, the molecules aren't the only ones.

References

1. Martyka, M., Tong, X.-Y., Jankowska, J., & Dral, P. O. (2026). OMNI-P2x universal neural network potential for excited-state simulations. Nature Communications. DOI: 10.1038/s41467-026-71380-5

2. Westermayr, J., & Marquetand, P. (2021). Machine Learning for Electronically Excited States of Molecules. Chemical Reviews, 121(16), 9873-9926. DOI: 10.1021/acs.chemrev.0c00749. PMID: 33211478

3. Li, Z., et al. (2025). Advancing nonadiabatic molecular dynamics simulations in solids with E(3) equivariant deep neural hamiltonians. Nature Communications, 16, 2138. DOI: 10.1038/s41467-025-57328-1

4. Axelrod, S., et al. (2022). Excited state non-adiabatic dynamics of large photoswitchable molecules using a chemically transferable machine learning potential. Nature Communications, 13, 3440. DOI: 10.1038/s41467-022-30999-w

5. Chen, W., et al. (2025). Machine learning for nonadiabatic molecular dynamics. Chemical Science. DOI: 10.1039/D5SC05579B

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.