Quantum ChemistryQuantum Chemistry

AIMNet2: The Neural Network Organic Chemistry Guide

AIMNet2 brings DFT-quality accuracy to routine organic chemistry calculations. A practical guide for medicinal chemists.

SciRouter Team
April 11, 2026
10 min read

AIMNet2 is one of the most practically useful neural network potentials for organic chemistry in 2026. It does not try to cover the whole periodic table. It focuses on the elements and chemistries that medicinal chemists actually work with, and within that scope it is very fast, very accurate, and very easy to integrate. This guide is a plain-English tour of what it does and when to use it.

What AIMNet2 is

AIMNet2 is the second generation of the AIMNet model family from the Isayev group at Carnegie Mellon. It is a neural network trained on a large dataset of density functional theory calculations covering drug-like organic molecules, and it learns to reproduce the DFT energies, forces, and partial atomic charges with high accuracy.

Unlike universal materials potentials, AIMNet2 is specialized. The training data emphasizes typical organic elements (H, C, N, O, F, P, S, Cl, Br, I) and the geometries and functional groups that appear in drug discovery. In exchange for narrower scope, you get better accuracy on the molecules you actually care about.

What AIMNet2 can compute

Because AIMNet2 produces energies, forces, and atomic charges, anything that reduces to those quantities is accessible. In practice that includes:

Conformer optimization and energies

Given a 3D structure, AIMNet2 will relax it to a local minimum at DFT-like accuracy in a fraction of a second. Generate 10 conformers, optimize each, compare energies, and pick the lowest. The whole cycle runs in seconds for drug-like molecules.

Bond dissociation energies

Compute the energy difference between a closed-shell molecule and the two radicals that would result from a bond break. AIMNet2 handles radicals, so it can give you BDE numbers directly. Useful for reasoning about metabolic stability, radical reactions, and oxidation.

Reaction barriers

With a nudged elastic band or string method on top of AIMNet2 forces, you can compute activation barriers for organic reactions in seconds per NEB image. This turns rough mechanistic thinking into something you can quantitatively explore.

pKa and logP via thermodynamic cycles

pKa is an energy difference between protonated and deprotonated forms in the relevant environment. logP is an energy difference between the molecule in octanol and in water. Both can be computed from the underlying AIMNet2 energy model with appropriate solvation corrections.

Atomic charges

The model outputs atomic charges along with the energy, which gives you a physically motivated electrostatic picture of the molecule. Useful for force field parameterization, docking grid generation, and understanding binding pockets.

When to use AIMNet2 versus other tools

Here is a rough decision tree for when AIMNet2 is the right pick:

  • Drug-like small molecule, need DFT-quality answers fast: AIMNet2 or Egret-1. See our Egret-1 overview for comparison.
  • Materials chemistry, broad element coverage: MACE-MP-0 or another universal materials potential.
  • Transition metal complex: DFT. Neither AIMNet2 nor MACE handles organometallics well.
  • Very large molecule or protein: A protein-specific or fragment-based approach, not a small-molecule potential.
Note
A useful pattern is to use AIMNet2 for the structural conformer optimization and Egret-1 for the derived property predictions. They are trained on similar data and give consistent numbers.

Speed in practice

On a single GPU, AIMNet2 evaluates a drug-like molecule in tens of milliseconds. On a CPU, it runs in a few hundred milliseconds. For an optimization that takes 30 steps to converge, that is one or two seconds total.

What this unlocks is screening at scale. A library of 100,000 candidates is a few hours on a modest GPU. You can compute conformer energies, atomic charges, or even full pKa distributions on every molecule in your virtual screening deck, rather than only the handful you have budget for with DFT.

Integration patterns

A few common ways to use AIMNet2 in production workflows:

  • Conformer filter upstream of docking. Generate many candidate conformers, rank by AIMNet2 energy, keep the lowest handful, and dock only those.
  • Protonation state assignment. Evaluate pKa for every ionizable site and assign the dominant microspecies at physiological pH before downstream modeling.
  • BDE-informed metabolic stability predictions. Rank C-H bonds by bond dissociation energy and flag molecules with unusually weak bonds as metabolically vulnerable.
  • Custom force field fits. Use AIMNet2 atomic charges and relaxed geometries as reference data for fitting molecular mechanics parameters.

Limitations worth knowing

AIMNet2 is trained on organic molecules, which means it does not generalize to unusual electronic structures, to metal-containing complexes, or to very large systems. If your target sits outside this envelope, AIMNet2 will silently produce numbers that look reasonable but may be quantitatively wrong. Spot-check with DFT if you are pushing the envelope.

Bottom line

AIMNet2 is a focused, practical tool for organic quantum chemistry. It is not universal. It is not a replacement for DFT for every workflow. But for drug discovery small molecules, it is often the fastest way to go from a SMILES string to a DFT-quality answer, and the easiest way to turn quantum chemistry into a routine screening step.

Try it in the QM properties tool →

Frequently Asked Questions

What is AIMNet2?

AIMNet2 is a neural network potential for organic chemistry from the Isayev group. It is trained on a large dataset of DFT calculations across drug-like molecules and delivers energies, forces, and atomic charges with accuracy close to the training DFT while running thousands of times faster.

How is AIMNet2 different from MACE?

MACE and MACE-MP-0 are universal materials potentials trained on broad inorganic datasets. AIMNet2 is specialized for organic chemistry with emphasis on drug-like molecules. The two are complementary, not competitive. Use AIMNet2 when you need the specific flavor of accuracy that organic medicinal chemistry demands.

What can I compute with AIMNet2?

Conformer optimization, relative conformer energies, bond dissociation energies, reaction barriers for common organic transformations, pKa and logP via thermodynamic cycles, and Hirshfeld charges for electrostatic analysis. Anything that reduces to energies, forces, and charges is accessible.

Does AIMNet2 handle tautomers?

Reasonably well for common tautomers. Because the model is explicit about atomic positions, it can distinguish between tautomeric forms if you prepare them as separate input structures. It does not enumerate tautomers automatically, so you need to generate the candidates yourself.

Is AIMNet2 fast enough for virtual screening?

Yes. A typical drug-like molecule runs in tens of milliseconds on a GPU and a few hundred milliseconds on a CPU. You can screen a library of 100,000 compounds for conformer energies or charge-based descriptors overnight on a single GPU.

Where can I try it?

AIMNet2 is available as part of the quantum chemistry properties stack on SciRouter, and the underlying model is also available as an open-source Python package from the Isayev group.

Try this yourself

500 free credits. No credit card required.