[Paper Review] Transferable Molecular Charge Assignment Using Deep Neural Networks

Ben Nebgen, Nick Lubbers+7 more|arXiv (Cornell University)|Mar 12, 2018

Machine Learning in Materials Science3 citations

TL;DR

This paper proposes a deep learning approach using the HIP-NN neural network architecture to predict atomic charges in molecules with high accuracy across diverse molecular structures and charge schemes. The method achieves quantum-level accuracy at ~104× faster speed than DFT, enabling efficient simulation of IR spectra from molecular dynamics trajectories while maintaining strong transferability across molecules and non-equilibrium geometries.

ABSTRACT

We use HIP-NN, a neural network architecture that excels at predicting molecular energies, to predict atomic charges. The charge predictions are accurate over a wide range of molecules (both small and large) and for a diverse set of charge assignment schemes. To demonstrate the power of charge prediction on non-equilibrium geometries, we use HIP-NN to generate IR spectra from dynamical trajectories on a variety of molecules. The results are in good agreement with reference IR spectra produced by traditional theoretical methods. Critically, for this application, HIP-NN charge predictions are about 104 times faster than direct DFT charge calculations. Thus, ML provides a pathway to greatly increase the range of feasible simulations while retaining quantum-level accuracy. In summary, our results provide further evidence that machine learning can replicate high-level quantum calculations at a tiny fraction of the computational cost.

Motivation & Objective

• To develop a machine learning model capable of accurately predicting atomic charges across a wide range of molecules and charge assignment schemes.
• To enable efficient simulation of molecular properties, such as IR spectra, by replacing computationally expensive DFT charge calculations.
• To demonstrate transferability of charge predictions across diverse molecular structures and non-equilibrium geometries.
• To achieve quantum-level accuracy in charge prediction at a fraction of the computational cost of traditional methods.

Proposed method

• The HIP-NN neural network architecture is trained to predict atomic charges directly from molecular geometry and atomic identities.
• The model is trained on a diverse dataset of molecules spanning small to large systems and various charge schemes.
• Charge predictions are used to compute dipole moments, which are then used to simulate IR spectra via time-dependent dipole response.
• The method is validated by comparing predicted IR spectra with reference spectra computed using traditional quantum chemistry methods.
• The computational efficiency is benchmarked by comparing HIP-NN inference time against direct DFT charge calculations.
• The model is evaluated on non-equilibrium geometries extracted from molecular dynamics trajectories to test robustness and transferability.

Experimental results

Research questions

• RQ1 Can a deep neural network accurately predict atomic charges across a broad range of molecular structures and charge schemes?
• RQ2 How does the performance of the HIP-NN model compare to DFT in predicting IR spectra from molecular dynamics trajectories?
• RQ3 To what extent is the charge prediction model transferable to non-equilibrium geometries and diverse molecular systems?
• RQ4 What is the computational speedup of the ML-based approach relative to direct DFT charge calculations?

Key findings

• The HIP-NN model achieves high accuracy in atomic charge prediction across diverse molecules, including both small and large systems.
• The predicted IR spectra from molecular dynamics trajectories show strong agreement with reference spectra computed using traditional theoretical methods.
• The HIP-NN-based charge prediction is approximately 10,000 times faster than direct DFT charge calculations, enabling large-scale simulations.
• The model demonstrates strong transferability to non-equilibrium geometries, maintaining accuracy without retraining.
• The method enables quantum-level accuracy in molecular property prediction at a drastically reduced computational cost.