ANI-1: An extensible neural network potential with DFT accuracy at force field computational cost (1610.08935v4)

Abstract: 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 fully transferable potential for organic molecules. We introduce ANAKIN-ME (Accurate NeurAl networK engINe for Molecular Energies) or ANI in short. ANI is a new method and procedure for training neural network potentials that utilizes a highly modified version of the Behler and Parrinello symmetry functions to build single-atom atomic environment vectors as a molecular representation. We utilize ANI to build a potential called ANI-1, which was trained on a subset of the GDB databases with up to 8 heavy atoms to predict total energies for organic molecules containing four atom types: H, C, N, and O. To obtain an accelerated but physically relevant sampling of molecular potential surfaces, we also propose a Normal Mode Sampling (NMS) method for generating molecular configurations. Through a series of case studies, we show that ANI-1 is chemically accurate compared to reference DFT calculations on much larger molecular systems (up to 54 atoms) than those included in the training data set, with root mean square errors as low as 0.56 kcal/mol.

• The paper introduces ANAKIN-ME (ANI), a novel machine learning method for constructing neural network potentials (NNPs) that achieve quantum mechanics-level accuracy at the computational cost of force field methods.
• ANI leverages modified Behler and Parrinello symmetry functions to create atomic environment vectors (AEVs) and is trained extensively on GDB-11 database subsets using a Normal Mode Sampling (NMS) strategy.
• The resulting ANI-1 potential demonstrates high accuracy with an RMSE of ~1.8 kcal/mol against DFT and shows significant transferability to larger organic molecules not included in the training set.

An Extensible Neural Network Potential for Organic Molecules: ANI-1

The presented work introduces a novel method for constructing neural network potentials (NNPs) with quantum mechanics-level accuracy at the computational cost of force field methods, termed ANAKIN-ME (ANI). This paper focuses on building a neural network potential, ANI-1, for organic molecules, reflecting a significant stride in increasing the accuracy and transferability of NNPs.

Central Thesis

The central thesis of this research is the development of ANI, a machine learning-based potential that leverages a modified set of Behler and Parrinello symmetry functions to create atomic environment vectors (AEVs). These vectors serve as molecular representations that significantly extend the transferability of NNPs to complex organic systems, demonstrated through the ANI-1 potential. The modification addresses key challenges in previous models, such as recognizing spatial atomic arrangements in different molecules and improving atom type differentiation.

Methodology

ANI's robustness emanates from extensive training using data derived from a subset of the GDB-11 database, specifically excluding fluorine, and focusing on molecules with up to 8 heavy atoms. A pivotal component of the method is the Normal Mode Sampling (NMS) strategy for generating conformational data, proving essential for spanning the molecular potential energy surface. Significant computational efforts were employed using NeuroChem software, leveraging GPU capabilities to optimize neural network training and atomic environment vector computations.

Key Results

The ANI-1 potential demonstrates substantial predictive accuracy, with a root mean square error (RMSE) of approximately 1.8 kcal/mol against reference density functional theory (DFT) calculations. ANI-1 outperforms popular semi-empirical methods such as PM6 and DFTB, particularly in predicting both absolute and relative conformational energies. This work extends the application of NNPs to larger molecular systems containing 10 to 54 atoms, showcasing its transferability beyond the training set's size.

In the test cases, ANI-1 maintains low energy prediction errors across structurally diverse organic compounds, highlighting its effectiveness in predicting isomer energetics, energy differences in retinol conformations, and potential energy scans of chemically relevant reactions. The methodology's adaptability to larger and chemically varied datasets is also noted as a key aspect driving the potential's effectiveness.

Implications and Future Directions

The implications of this research are manifold. The ANI-1 potential offers a tool with potential applications in molecular dynamics and optimization tasks, effectively bridging the computational cost gap between classical force fields and quantum mechanical simulations. Looking forward, augmenting the ANI-1 dataset with additional molecules and atom types could further refine its accuracy and broaden its applicability to other chemical environments.

This work paves the way for future advancements in creating more extensive and chemically accurate NNPs that could be tailored for different classes beyond organic molecules, enhancing the toolset available for computational chemists and material scientists.

In conclusion, the development of ANI-1 represents a noteworthy contribution to the field of computational chemistry, demonstrating the power of machine learning in developing transferable, accurate, and computationally efficient models of molecular interactions. As computational resources and dataset accuracy improve, the ANI methodology can be expected to play a significant role in advancing the theoretical paper of chemical systems.