Source-linked AI summary
SchNet: A continuous-filter convolutional neural network for modeling quantum interactions
Kristof T. Schütt, Pieter-Jan Kindermans, Huziel E. Sauceda, Stefan Chmiela, Alexandre Tkatchenko, Klaus-Robert Müller
TL;DR
Quantum-chemical machine learning must model atomistic systems beyond equilibrium structures and avoid losing positional information through grid discretization. The paper introduces continuous-filter convolutions in SchNet to predict energies and forces while respecting quantum-chemical constraints, achieving strong results on equilibrium and molecular-dynamics benchmarks but leaving combined chemical and structural variation difficult.
Problem
Chemical-space exploration is constrained by the vastness of possible compounds, costly accurate quantum-chemical calculations, and datasets focused mainly on equilibrium molecules.
Method
SchNet uses continuous-filter convolutions to model atoms at arbitrary positions and jointly predicts energies and forces with quantum-chemical invariances, smoothness, and energy conservation.
Results
SchNet improves state-of-the-art equilibrium-energy prediction, accurately predicts energies and forces across MD17 trajectories, and introduces ISO17 for chemical and conformational variation.
Takeaways & Limitations
Modeling both chemical and conformational variations is an important step toward machine-learning-driven quantum-chemical exploration, although ISO17 remains difficult.
Takeaways & Limitations
Force-based training requires about twice the computation time because obtaining forces requires full forward and backward passes through the energy model.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper targets machine-learning models that handle both chemical and structural variation while preserving quantum-chemical constraints. It introduces continuous-filter convolutions and SchNet to predict energies and forces for atomistic systems.
- Chemical-space exploration is limited by the vastness of possible compounds and the cost of accurate quantum-chemical calculations.
- The model must represent equilibrium and non-equilibrium conformations while producing smooth, energy-conserving force fields with appropriate molecular invariances.
- Continuous-filter convolutional layers extend convolution beyond grid-bound data to objects with arbitrary positions, including atoms in molecules and materials.
- SchNet predicts rotationally invariant energies and rotationally equivariant forces using a smooth potential energy surface and an energy-conserving force field.
- ISO17 benchmarks chemical and conformational changes, and training with forces improves generalization in this setting.
2 Related work
Prior approaches include feature-based models, molecular graph methods, GDML, DTNN, and enn-s2s, each with limitations in scalability, continuity, or molecular-dynamics applicability. SchNet addresses discontinuous energy surfaces associated with discrete input features.
- Earlier neural-network and Gaussian-process approaches fit potential energy surfaces using hand-crafted features, while graph models encode neighborhood and bond relationships.
- GDML enforces energy conservation by learning energy gradients directly, but its kernel matrix grows quadratically with atom and example counts.
- DTNN shares SchNet’s invariances but lacks the continuous-filter convolution interpretation and falls behind SchNet and enn-s2s in accuracy.
- enn-s2s achieves state-of-the-art QM9 performance but cannot predict molecular dynamics because discrete input encodings create discontinuities in its potential energy surface.
- SchNet instead uses continuous-filter convolutions to produce a continuous potential energy surface without one-hot input features.
3 Continuous-filter convolutions
Continuous-filter convolutions replace grid-defined filters with position-dependent filters generated for unevenly spaced inputs. This lets SchNet model interactions among atoms at arbitrary positions while handling feature interactions separately.
- Standard convolutions assume discretized signals, but molecular atoms occupy unevenly spaced positions that cannot be represented naturally on a shared grid.
- The continuous-filter approach uses a neural network to generate filter values from positions, generalizing discrete filter tensors to arbitrary positions and object counts.
- The convolutions are applied feature-wise for efficiency, while separate atom-wise layers handle interactions between feature maps.
4 SchNet
SchNet represents molecules atom-wise and updates those representations through geometry-dependent interaction blocks. Its continuous-filter convolutions use distances to model local interactions while preserving smooth, physically constrained energy and force predictions.
- Architecture: SchNet represents each molecule atom-wise, then pools updated atom-wise energies for the final prediction.The architecture uses atomic features initialized from atom types and updates them through interaction blocks.
- Architecture: Shared atom-wise layers recombine feature maps independently for each atom, supporting scalability across molecular sizes.The interaction blocks keep 64 feature maps and use residual updates containing atom-wise and interatomic layers.
- Continuous-filter convolution: Continuous-filter convolutions generate rotationally invariant filters from interatomic distances and expand distances with radial basis functions.Radial basis expansion reduces initial filter correlations and supports a faster training procedure.
- Continuous-filter convolution: Three interaction blocks emphasize different distance ranges, enabling sequential construction of complex many-body representations while retaining rotational invariance.The filters are radial and their updates depend on each atom's radial environment.
- Energy and forces: Forces are obtained by differentiating the predicted energy with respect to atomic positions, making the force field energy-conserving and rotationally equivariant.The shifted softplus is used throughout to obtain a smooth potential energy surface, while joint energy-force training combines both targets in the loss.
5 Experiments and results
The experiments evaluate SchNet across equilibrium molecules, molecular dynamics trajectories, and chemically and structurally varied isomers. Performance is strong on QM9 and MD17, while ISO17 remains challenging and force training improves generalization.
- QM9 – chemical degrees of freedom: SchNet consistently achieves state-of-the-art QM9 energy prediction, reaching an MAE of 0.31 kcal/mol with 110k training examples.
- MD17 – conformational degrees of freedom: MD17 evaluates separate energy-and-force models on eight single-molecule trajectories spanning diverse conformations.
- MD17 – conformational degrees of freedom: Including forces gives SchNet accurate MD17 energies and forces, with errors below 0.12 kcal/mol and 0.33 kcal/mol/Å.
- MD17 – conformational degrees of freedom: With 50,000 energy-only training examples, SchNet outperforms DTNN on three of four MD17 datasets, demonstrating its scalability.
- ISO17 – chemical and conformational degrees of freedom: ISO17 combines chemical and conformational changes using short molecular-dynamics trajectories of 129 isomers and tests both known and unseen molecules.
- ISO17 – chemical and conformational degrees of freedom: On ISO17, force training improves performance for both known and unseen molecules, but chemical and structural variation remains difficult.
6 Conclusions
The paper introduces continuous-filter convolutional layers and SchNet for modeling quantum interactions on unevenly spaced atomic data. Across increasingly difficult benchmarks, SchNet improves equilibrium-energy prediction and accurately models molecular-dynamics energies and forces, while ISO17 exposes remaining challenges.
- Continuous-filter convolutional layers model unevenly spaced data, including atoms in molecules and materials.
- SchNet respects quantum-chemical constraints through rotationally invariant energies and rotationally equivariant, energy-conserving force predictions.
- SchNet improves state-of-the-art equilibrium energy prediction on QM9 and accurately predicts energies and forces across MD17 trajectories.
- ISO17 provides a challenging benchmark with 645,000 conformations and shows that chemical and conformational variations still need further improvement.