SchNet: A continuous-filter convolutional neural network for modeling quantum interactions
By Kristof T. Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, K. Müller. 2017
1,449 citations recorded by the source corpus.
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.
Subfields: Artificial Intelligence, Materials Chemistry, Physical and Theoretical Chemistry, Model Architectures, AI for Physical Sciences
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Yoshua Bengio
- Jianfeng Gao
- Percy Liang
- Mohit Bansal
- Jure Leskovec
- Dacheng Tao
- Quoc V. Le
- William L. Hamilton
In Materials Chemistry
- Tess E. Smidt
- Tommi S. Jaakkola
- Regina Barzilay
- Jian Tang
- Shitong Luo
- Frank No'e
- K. Müller
- Max Welling
In Physical and Theoretical Chemistry
- Tess E. Smidt
- Tommi S. Jaakkola
- Risi Kondor
- Regina Barzilay
- Shitong Luo
- Jian Tang
- Evgeny V. Podryabinkin
- Wengong Jin
In Model Architectures
- Tess E. Smidt
- Tommi S. Jaakkola
- Max Welling
- Regina Barzilay
- Erik J. Bekkers
- Risi Kondor
- Daniel E. Worrall
- Mario Geiger
In AI for Physical Sciences
- Risi Kondor
- Frank No'e
- Tess E. Smidt
- K. Müller
- Grgoire Montavon
- Shitong Luo
- Stephan Günnemann
- Johannes Klicpera
Citations in the corpus
- E(n) Equivariant Graph Neural Networks (2021, methodology): “Schnet (Sch¨utt et al., 2017b) can be interpreted as an E( n ) invariant Graph Neural Network where φ cf receives as input relative distances and outputs a continuous filter convolution that multiplies the neighbor embeddings h .”
- Directional Message Passing for Molecular Graphs (2020, methodology): “We use 6 state-of-the-art models for comparison: SchNet (Schütt et al., 2017), PhysNet (results based on the reference implementation) (Unke & Meuwly, 2019), provably powerful graph networks (PPGN, results provided by the original authors) (Maron et al., 2019), MEGNet-simple (without auxiliary…”
- Hierarchical Graph Representation Learning with Differentiable Pooling (2018, methodology): “GNNs have been applied to a wide variety of tasks, including node classification [16, 21], link prediction [31], graph classification [7, 11, 39], and chemoinformatics [14, 19, 27, 28, 32].”
- Learning Multimodal Graph-to-Graph Translation for Molecular Optimization (2018, neutral): “, 2017) and chemistry (Kearnes et al., 2016; Gilmer et al., 2017; Schütt et al., 2017; Jin et al., 2017).”
- Junction Tree Variational Autoencoder for Molecular Graph Generation (2018, methodology): “For applications, graph neural networks are used in semisupervised classification (Kipf & Welling, 2016), computer vision (Monti et al., 2016), and chemical domains (Kearnes et al., 2016; Schütt et al., 2017; Jin et al., 2017).”
- SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks (2020, background): “[19], output channels are a learned linear combination of input channels using one set of weights w i,c′c = w `` c′c per representation degree, shared across all points.”
- Equivariant Flows: exact likelihood generative learning for symmetric densities (2020, methodology): “Other Invariant Potential Functions While Φ could be modeled by any kind of invariant graph neural networks, such as SchNet (Schütt et al., 2017), this would require us to 1) use AD in order to compute ∇x(t)Φ(x(t)) and 2) compute ∆x(t)Φ(x(t)) at every function evaluation while integrating v.”
- Cormorant: Covariant Molecular Neural Networks (2019, methodology): “To the best of our knowledge, the current state-of-the art algorithms on this dataset are DeepMD [Zhang et al., 2018], DTNN [Schütt et al., 2017], SchNet [Schütt et al., 2017], GDML [Chmiela et al., 2016], and sGDML [Chmiela et al., 2018].”
- Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds (2018, methodology): “Many of these use only the pairwise distance between atoms (e.g. SchNet [13] and the graph convolutional model from Faber et al. [14]).”
- Neural Message Passing with Edge Updates for Predicting Properties of Molecules and Materials (2018, methodology): “A range of graph convolution models (Duvenaud et al., 2015; Li et al., 2015; Battaglia et al., 2016; Kearnes et al., 2016; Schütt et al., 2017a) including Laplacian based models (Bruna et al., 2013) can be cast into this message passing framework as shown by Gilmer et al. (2017).”
- TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction (2022, background): “These works have been shown great potential in various applications like protein structure modeling [13, 17–19], molecular low-energy generation prediction [24–26], property/function prediction [22, 23] and molecule design [20, 21].”
- DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking (2022, methodology): “These are concatenated with sinusoidal embeddings of the diffusion time [21] and, in the case of edges, radial basis embeddings of edge length [22].”
- SplineCNN: Fast Geometric Deep Learning with Continuous B-Spline Kernels (2017, background): “ Therefore, our approach is orthogonal to improvements in this field. Spatial continuous convolution kernels. While the first continuous convolution kernels for graphs work in the spectral domain (e.g. [9, 6, 20]), spatial continuous convolution kernels for irregular structured data were introduced re”
- Swin Transformer V2: Scaling Up Capacity and Resolution (2021, methodology): “Our Log-CPB approach is also related to earlier works on continuous convolution and variants [30, 45, 58, 67], which utilize a meta network to handle irregular data points.”