SchNet - a deep learning architecture for molecules and materials (1712.06113v3)

Abstract: Deep learning has led to a paradigm shift in artificial intelligence, including web, text and image search, speech recognition, as well as bioinformatics, with growing impact in chemical physics. Machine learning in general and deep learning in particular is ideally suited for representing quantum-mechanical interactions, enabling to model nonlinear potential-energy surfaces or enhancing the exploration of chemical compound space. Here we present the deep learning architecture SchNet that is specifically designed to model atomistic systems by making use of continuous-filter convolutional layers. We demonstrate the capabilities of SchNet by accurately predicting a range of properties across chemical space for \emph{molecules and materials} where our model learns chemically plausible embeddings of atom types across the periodic table. Finally, we employ SchNet to predict potential-energy surfaces and energy-conserving force fields for molecular dynamics simulations of small molecules and perform an exemplary study of the quantum-mechanical properties of C$_{20}$-fullerene that would have been infeasible with regular ab initio molecular dynamics.

• The paper introduces SchNet, a deep learning model that uses continuous-filter convolutions to accurately capture atomic interactions and predict molecular and material properties.
• It details a methodology incorporating atom-wise layers and interaction blocks that embed rotational and translational invariance into atomistic representations.
• Empirical evaluations on datasets like QM9 and the Materials Project demonstrate SchNet’s superior precision, promising efficient simulations and materials discovery.

An Overview of SchNet: A Deep Learning Architecture for Molecules and Materials

The paper presents SchNet, a novel deep learning architecture specifically tailored for the modeling of atomistic systems, focusing on molecules and materials. This paper addresses the computational challenges in chemical physics, particularly the costly nature of quantum-mechanical calculations and the need to explore vast chemical compound spaces efficiently.

SchNet leverages continuous-filter convolutional layers to learn representations of atomistic systems from scratch, respecting fundamental symmetries such as rotational and translational invariance. This ability to embed symmetries directly into the architecture allows for modeling complex atomic interactions essential for predicting potential-energy surfaces and carrying out molecular dynamics simulations.

Methodology

SchNet builds on earlier work with Deep Tensor Neural Networks (DTNN) but introduces significant advancements. It uses:

• Atom-wise layers: Each atom in the system is represented in layers that are updated based on interactions with neighboring atoms.
• Interaction blocks: These employ continuous-filter convolutional (cfconv) layers, which extend traditional discrete convolutions to arbitrary atom positions, crucial for capturing atomic interactions accurately.
• Filter-generating networks: These networks model the interactions using either fixed or periodic boundary conditions, capturing the nuanced effects of the atomic environment.

Results and Implications

The performance of SchNet was evaluated on the QM9 dataset, predicting molecular properties and achieving mean absolute errors that often surpass competitive models like enn-s2s in terms of accuracy for certain properties. SchNet also demonstrated the capability to predict formation energies of bulk crystals within the Materials Project dataset with high precision.

One significant outcome is the capacity of SchNet to learn chemically plausible atom type embeddings, indicating that without explicit programming, the model can capture periodic trends across different chemical spaces. This machine learned understanding was notably demonstrated through local chemical potentials, a novel interpretative tool offered by SchNet, which visualizes the chemical environment surrounding atoms in molecules and materials.

Additionally, SchNet's application to molecular dynamics, specifically for the C$_{20}$ fullerene at the PBE+vdW$^{\rm TS}$ level of theory, illustrates its utility in executing simulations that otherwise would be computationally prohibitive with conventional ab initio methods.

Future Directions

The paper suggests avenues for further research and development by increasing the interpretability of neural network models, thus enhancing their utility not just as predictive tools but also as a means for gaining chemical insight. Efforts may involve refining property-specific output layers or integrating domain-specific knowledge directly into the architecture to further improve model performance and interpretability.

Furthermore, the scalability of SchNet to larger datasets and more complex systems opens the door to its application in real-world material science and pharmaceutical research, potentially accelerating discovery and innovation.

Conclusion

SchNet represents a sophisticated integration of deep learning into quantum chemistry, providing a robust framework for addressing computational challenges in modeling molecular and material properties. It showcases how machine learning, particularly deep architectures, can adjust dynamically to capture the underlying physics of atomic interactions, highlighting its promising role in the future of computational chemistry and materials science research.