Quantum chemistry can be used to calculate the probabilities of which chemical processes will occur. This involves a great deal of programming work. Dr. Michael F. Peintinger from the Mulliken Center for Theoretical Chemistry at the University of Bonn has investigated what happens when artificial intelligence (AI) is used for this purpose.
What is it all about?
Almost all of a substance’s properties depend on how its electrons are distributed and how they behave: whether a material conducts electricity, what color a dye has, how stable a crystal is, and how fast a reaction proceeds. This is called the electronic structure.
Do you have a specific example?
For example, I’m interested in how a molecule adsorbs onto a solid surface—such as water on a wafer-thin layer of boron nitride, a material made of boron and nitrogen. Where does a water molecule prefer to sit on this layer, and how firmly does it adhere to the surface there? To find that out, I would place the molecule at different locations in the model, orient it in different ways, and compare the energies. The arrangement with the lowest energy is the most favorable. Understanding how such molecules adsorb onto surfaces is particularly important for understanding catalytic reactions.
What is quantum chemistry?
Quantum chemistry applies the laws of quantum mechanics to chemical problems. Electrons do not move along fixed paths. Instead, we chemists calculate the probability that an electron will be found at a specific location. In practice, the equations can be solved exactly and analytically only for a single electron, such as in a hydrogen atom. Even in a small piece of diamond, many electrons influence one another. Chemists therefore need approximations—and which one is appropriate depends on the question at hand.
Can you use this to predict chemical reactions?
To some extent. For a specific reaction, we can calculate how much energy is released or must be supplied. We can also trace a possible reaction pathway. In the process, bonds break and new ones form, and in most cases, an energy barrier must first be overcome. The height of this barrier determines how fast the reaction proceeds at a given temperature. The question “What happens when you mix these two substances?” is more difficult. To answer that, you need to know which reaction pathways compete with one another, what role solvents and temperature play, and whether a catalyst is involved. A calculation alone can rarely answer that.
There are many quantum chemistry programs. What’s new about vibe-qc?
To the best of my knowledge, vibe-qc is the first program of its kind to be developed almost entirely by generative AI. So what’s new is the way it works, not a single computational method. I designed a system in which AI agents—a kind of digital programmer—write and test a program of this size under scientific guidance, and they do so very quickly: The first comprehensive version was ready after about ten weeks.
How do you ensure that the AI does what it’s supposed to?
I have to review every suggestion. On the first day, the assistant wanted to calculate a lithium atom using a method that treats all electrons in pairs. But lithium has three electrons, so one remains unpaired. That sounded plausible, but it was easy to see that it was wrong just by counting. Other errors weren’t noticed until the results were compared with those of an independently developed program. That’s why a method in vibe-qc is only considered ready for use once it stands up to such comparisons.
How close is the calculation to reality?
How well a calculation describes nature also depends on how much the model simplifies reality. This became evident when considering how much energy a chemical reaction needs to get started. Here, the results from the simplified model deviated significantly, on average, from those of more complex calculations. Both sets of results came from the computer. To verify how close a prediction is to reality, I need measured values from the lab.
What do you consider the most important result?
For me, the most important result is that this approach works. Within just a few months, vibe-qc has become a general-purpose program for molecules and crystals. It supports many common computational methods, ranging from fast approximations to very complex, precise methods. Today, a single researcher can run a program of this scale instead of writing it from scratch. vibe-qc and its associated programs are open source. The source code is publicly available, and anyone is free to use, modify, and distribute it under the Mozilla Public License.
What was the biggest challenge?
A program can run without any error messages and produce a number that looks convincing. It becomes difficult when a second program yields a different number. In a comparative calculation, for example, the two programs had treated the inner electrons of the atoms differently and therefore did not include the same electrons in the calculation. Both numbers were correct but pertained to different questions. To clarify something like this, you have to go back to the underlying physical assumptions.
Is there already an application?
So far, the published calculations have primarily served as tests. For diamond, for example, I calculated the energies that electrons in the crystal can assume and, for a selected case, examined the magnitude of the energy jump from an occupied to a free state. An independent calculation using a different program yielded almost the same value.
What is the source?
Michael F. Peintinger: VIBE-QC: An Open-Source, General-Purpose Electronic-Structure Code for Molecules and Solids Developed With Generative AI, Journal of Computational Chemistry, https://onlinelibrary.wiley.com/doi/10.1002/jcc.70525
Where can I find out more?
Dr. Michael Peintinger, Mulliken Center for Theoretical Chemistry at the University of Bonn, Tel. +49 228 7360443, Email: mpei@uni-bonn.de
Journal of Computational Chemistry
Computational simulation/modeling
Not applicable
VIBE-QC: An Open-Source, General-Purpose Electronic-Structure Code for Molecules and Solids Developed With Generative AI
27-Oct-2026