Ewald-Based Long-Range Message Passing for Molecular Graphs

Explore an in-depth conference talk on Ewald-Based Long-Range Message Passing for Molecular Graphs presented by Arthur Kosmala from Valence Labs. Dive into the world of neural architectures for learning potential energy surfaces from molecular data, focusing on the Message Passing Neural Network (MPNN) paradigm. Discover how the proposed Ewald message passing technique addresses the limitations of traditional MPNNs by incorporating long-range interactions such as electrostatics and van der Waals forces. Learn about the theoretical foundations of this approach in the Ewald summation method and its implementation as an augmentation to existing MPNN architectures. Examine the results of experiments conducted on diverse periodic and aperiodic structures, showcasing improvements in energy mean absolute errors across multiple models and datasets. Gain insights into the impact of this technique on structures with high long-range contributions to ground truth energy. The talk covers various topics including problem setting, MPNN blueprint, continuous-filter convolutions, periodic boundary conditions, Ewald summary, long-range message sums, the aperiodic case, radial vs. non-radial filtering, combination with existing models, baselines and datasets, comparison studies, analysis of long-range impact, and force vs. energy results.

- Intro
- Problem Setting
- MPNN Blueprint
- Continous-Filter Convolutions
- Periodic Boundary Conditions
- Ewald Summary
- Ewald Message Passing
- Long-Range Message Sums
- The Aperiodic Case
- Radial vs. Non-Radial Filtering
- Combination with Existing Models
- Baselines & Datasets
- Comparison Studies
- Analysis of Long-Range Impact
- Force vs. Energy Results
- Conclusion
- Q+A