Open Catalyst Project (OC20)

The core task is to predict the potential energy and per-atom forces of a catalyst-adsorbate system directly from the 3D atomic positions. This is a regression problem that requires models to learn the quantum mechanical potential energy surface. Success is measured by Mean Absolute Error (MAE) on energy and forces, with a focus on force prediction as it drives structural relaxation.

• Input: 3D atomic coordinates and atomic numbers
• Output: System energy (eV) and per-atom force vectors (eV/Å)
• Primary metric: Force MAE within a threshold

A general framework for graph neural networks where node representations are iteratively updated by aggregating information from neighboring nodes via message and update functions. In the context of OC20, MPNNs form the backbone of models that learn to predict energy from atomic structures.

• Message Function: Computes a message from a source atom to a target atom based on their features and distance.
• Update Function: Integrates aggregated messages to produce a new node representation.
• Readout Phase: Pools all node features into a single graph-level output, such as the system's adsorption energy.

A property of a function or neural network ensuring that its output transforms consistently with any input rotation and translation in 3D Euclidean space. This is critical for modeling molecular geometries because physical properties like energy are invariant to rotation, while forces are equivariant.

• Invariance: The predicted energy of a molecule does not change if the entire system is rotated.
• Equivariance: The predicted force vectors rotate identically with the molecule.
• Implementation: Achieved through spherical harmonics, tensor products, and Clebsch-Gordan coefficients in architectures like NequIP.

A GNN architecture that directly learns material properties from crystal structures by constructing a multigraph representing atomic connectivity across periodic boundary conditions. It was a strong baseline for early OC20 tasks.

• Periodic Multi-Graph: Atoms are connected to their nearest neighbors within a cutoff radius, accounting for the repeating unit cell.
• Edge Features: Interatomic distances are expanded into a radial basis to form the initial edge features.
• Property Prediction: After message passing, a pooling layer aggregates atom vectors to predict bulk material properties like formation energy.

A machine-learned surrogate model that predicts the potential energy and atomic forces of a molecular system directly from atomic coordinates, bypassing the explicit solution of the Schrödinger equation. The OC20 dataset was created using Density Functional Theory (DFT) to train such NNPs.

• Speed-Accuracy Trade-off: NNPs are orders of magnitude faster than DFT while aiming to retain quantum mechanical accuracy.
• Training Data: Requires large datasets of DFT-calculated energies and forces, like OC20.
• Force Prediction: A key requirement is that forces are the negative gradient of the predicted energy with respect to atomic positions.

A computationally efficient architecture that achieves E(n) equivariance without requiring expensive higher-order tensor products. It operates directly on scalar and vector features, making it a practical choice for large-scale molecular dynamics.

• Scalar Features: Invariant node embeddings (e.g., atom type encoding).
• Vector Features: Equivariant directional information that rotates with the molecule.
• Message Passing: Updates vector features using a weighted sum of relative position vectors, ensuring equivariance without Clebsch-Gordan decompositions.

A pioneering continuous-filter convolutional neural network that models quantum interactions by using interatomic distances to generate filter kernels for message passing. It established a foundational architecture for learning molecular properties.

• Continuous Convolution: Filter-generating networks map interatomic distances to smooth, continuous filters.
• Energy Conservation: By predicting energy as a sum of atom-wise contributions, SchNet naturally provides conservative force predictions.
• Interaction Blocks: Stacked modules of atom-wise updates, continuous convolutions, and non-linearities to capture many-body effects.