[Paper Review] SchNet: A continuous-filter convolutional neural network for modeling quantum interactions

Kristof T. Schütt, Pieter-Jan Kindermans+4 more|arXiv (Cornell University)|Jun 26, 2017

Machine Learning in Materials Science31 references470 citations

TL;DR

SchNet introduces continuous-filter convolutional layers to model quantum interactions in molecules, enabling energy predictions and energy-conserving force fields that respect quantum-chemical principles, achieving state-of-the-art results on QM9, MD17, and ISO17 datasets.

ABSTRACT

Deep learning has the potential to revolutionize quantum chemistry as it is ideally suited to learn representations for structured data and speed up the exploration of chemical space. While convolutional neural networks have proven to be the first choice for images, audio and video data, the atoms in molecules are not restricted to a grid. Instead, their precise locations contain essential physical information, that would get lost if discretized. Thus, we propose to use continuous-filter convolutional layers to be able to model local correlations without requiring the data to lie on a grid. We apply those layers in SchNet: a novel deep learning architecture modeling quantum interactions in molecules. We obtain a joint model for the total energy and interatomic forces that follows fundamental quantum-chemical principles. This includes rotationally invariant energy predictions and a smooth, differentiable potential energy surface. Our architecture achieves state-of-the-art performance for benchmarks of equilibrium molecules and molecular dynamics trajectories. Finally, we introduce a more challenging benchmark with chemical and structural variations that suggests the path for further work.

Motivation & Objective

• Develop a neural network that learns energy and force fields for molecules while respecting rotational/translation invariance and energy conservation.
• Extend neural networks to non-grid, atom-centered data using continuous-filter convolutions.
• Demonstrate the approach on benchmarks with chemical and conformational variation (QM9, MD17, ISO17).
• Show that including forces in training improves generalization across chemical compound space.

Proposed method

• Introduce continuous-filter convolutional (cfconv) layers that generate filters from interatomic distances via a filter-generating network.
• Represent molecules atom-wise with shared atom-type embeddings and apply interaction blocks that update atomic representations through cfconv and atom-wise layers.
• Use radial basis expansion of interatomic distances to ensure rotational invariance of filters.
• Train to predict total energy and atomic forces by minimizing a combined loss that includes energy and force terms, ensuring energy conservation.
• Guarantee smooth, curl-free force fields by construction, yielding rotationally equivariant force predictions.
• Fix network depth and feature map size (F=64) for the interaction blocks, with residual connections to enable complex many-body representations.

Experimental results

Research questions

• RQ1 Can SchNet achieve state-of-the-art energy predictions on equilibrium molecular datasets (QM9)?
• RQ2 How well does SchNet predict energies and forces for molecular dynamics trajectories (MD17) and for chemically diverse, structurally varied data (ISO17)?
• RQ3 Does training with forces improve generalization across chemical space compared to training on energies alone?
• RQ4 Are the learned energy surfaces smooth and energy-conserving, with rotational invariance and translation invariance preserved?
• RQ5 Can a single unified model represent both chemical and conformational variations across multiple molecules?

Key findings

• SchNet achieves state-of-the-art mean absolute error for QM9 energy predictions, reaching 0.31 kcal/mol at 110k training examples.
• On MD17, SchNet with energies alone and with energies plus forces yields competitive or superior energy and force predictions compared to GDML and DTNN across several trajectories, especially with larger training sets.
• Training with forces substantially improves predictive performance on MD17, particularly for flexible molecules, and SchNet remains scalable to larger molecules and datasets.
• On ISO17, SchNet attains chemical accuracy for known molecules and shows improved generalization when trained with forces, handling both chemical and conformational variations.
• Across QM9, MD17, and ISO17, SchNet produces smooth, energy-conserving force fields and rotationally invariant energy predictions, demonstrating the model’s adherence to fundamental quantum-chemical principles.