± SCIENCE Today’s artificial intelligence still faces two major challenges. One is that, in most industries, data exist…+ SCIENCE Deep learning takes on protein folding In 1972, Anfinsen won a Nobel prize for demonstrating a connection bet…± SCIENCE Significance Deep neural networks are currently the most successful machine-learning technique for solving a…+ SCIENCE Deep learning allows computational models that are composed of multiple processing layers to learn representa…+ HEALTH For the task of analyzing survival data to derive risk factors associated with mortality, physicians, researc…+ SCIENCE Artificial intelligence (AI) coupled with promising machine learning (ML) techniques well known from computer…+ SCIENCE Deep learning is revolutionizing many areas of science and technology, especially image, text, and speech rec…+ SCIENCE Current deep-learning models are mostly built upon neural networks, i.e. multiple layers of parameterized dif…
TruaceTracing the truth around AIMonday, July 13, 2026
Subscribe
Science·G Space·Evidence-backed gain·Published 2026-07-13

ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost

Deep learning is revolutionizing many areas of science and technology, especially image, text, and speech recognition. In this paper, we demonstrate how a deep neural network (NN) trained on quantum mechanical (QM) DFT calculations can learn an accurate and transferable potential for organic molecules. We introduce ANAKIN-ME (Accurate NeurAl networK engINe for Molecular Energies) or ANI for short. ANI is a new method designed with the intent of developing transferable neural network potentials that utilize a hig…

TRV-2026-0206Peer-reviewedPermanent record — cite & verify
ANI-1: an extensible neural network potential with DFT accuracy at force field computational cost

Atomic-view-of-the-histidine-environment-stabilizing-higher-pH-conformations-of-pH-dependent-ncomms8771-s2 by Valéry C, Deville-Foillard S, Lefebvre C, Taberner N, Legrand P, Meneau F, Meriadec C, Delvaux C, Bizien T, Kasotakis E, Lopez-Iglesias C, Gall A, Bressanelli S, Le Du M, Paternostre M, Artzner F. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

The quick read

On Feb 8 2017, researchers described ANI, a deep neural network architecture that uses atomic environment vectors derived from modified symmetry functions to learn molecular energies from quantum mechanical DFT calculations, and introduced Normal Mode Sampling to efficiently sample potential surfaces.

The work matters because it demonstrated a path to replace expensive DFT calculations with a transferable neural network potential that retains chemical accuracy while running at force-field cost, though by the publication date validation was limited to H, C, N, O chemistry and generalization from small training molecules to larger ones.

Main points
  • Introduces ANAKIN-ME (ANI) using modified Behler and Parrinello symmetry functions to build single-atom atomic environment vectors (AEV) as molecular representation.
  • Proposes Normal Mode Sampling (NMS) method for accelerated but physically relevant sampling of molecular potential surfaces to generate conformations.
  • ANI-1 trained on subset of GDB databases with up to 8 heavy atoms for H, C, N, O and tested on larger systems up to 54 atoms.
Gain

A deep neural network trained on DFT data learned a transferable potential that predicts total energies for organic molecules with chemical accuracy versus DFT, generalizing to systems up to 54 atoms despite training on smaller molecules.

The rundown

The authors built ANI using highly-modified Behler and Parrinello symmetry functions to create atomic environment vectors that span configurational and conformational space, and introduced Normal Mode Sampling to generate training conformations.

Case studies reported in the paper show ANI-1 maintains chemical accuracy versus reference DFT on systems much larger than the training set, up to 54 atoms, after training only on GDB molecules with up to 8 heavy atoms of H, C, N, O.

Sources

Reader signal

How should this claim be treated?

The debate