SOAP Descriptor

A systematic and complete basis set expansion of atomic environments that yields highly efficient, body-ordered invariant features. ACE provides a mathematically rigorous framework for constructing interatomic potentials, offering a direct theoretical connection to the SOAP descriptor. While SOAP uses a fixed Gaussian smearing and radial basis, ACE generalizes this to a complete, hierarchical expansion that can be systematically converged. This makes ACE both a powerful standalone method and a theoretical lens through which to understand the completeness of SOAP-like representations.

A fundamental symmetry property ensuring that a model's output transforms consistently with any input rotation and translation in 3D Euclidean space. SOAP descriptors are explicitly invariant to these transformations, making them ideal inputs for conventional ML models. Modern architectures like NequIP and MACE build equivariance directly into the neural network layers, learning more complex functions of atomic coordinates while preserving physical symmetries. Understanding this distinction between invariant features and equivariant operations is critical for designing models that respect molecular geometry.

A machine-learned surrogate model that predicts the potential energy and atomic forces of a molecular system directly from atomic coordinates. SOAP descriptors serve as a foundational input representation for many classical NNPs, including the pioneering GAP (Gaussian Approximation Potential) framework. By converting 3D coordinates into a smooth, rotationally invariant feature vector, SOAP enables kernel-based or neural network regression to learn the Born-Oppenheimer potential energy surface with high fidelity, bypassing expensive quantum mechanical calculations.

A fixed-length bit vector encoding the presence or absence of specific substructural features within a molecule. Unlike SOAP, which captures continuous 3D geometry, traditional fingerprints like ECFP or MACCS keys operate on the 2D molecular graph. SOAP can be seen as a 3D generalization, providing a smooth, differentiable, and geometrically aware alternative. While 2D fingerprints are faster to compute and useful for property prediction, SOAP's 3D sensitivity makes it essential for tasks where conformation and spatial arrangement dictate function.

A pioneering continuous-filter convolutional neural network that models quantum interactions using interatomic distances. SchNet learns its own atomic environment representations directly from data, effectively discovering features analogous to SOAP but optimized end-to-end for a specific task. Comparing SOAP (a fixed, physics-based descriptor) with SchNet's learned filters highlights a key design choice in computational chemistry: the trade-off between the interpretability and transferability of hand-crafted features versus the flexibility of learned representations.

A GNN architecture that directly learns material properties from crystal structures by constructing a multigraph representing atomic connectivity across periodic boundary conditions. While SOAP provides a local descriptor of atomic environments, CGCNN operates on the entire periodic graph. SOAP features can be used as input node features for CGCNN, combining the physics-informed representation of local geometry with the relational learning capabilities of graph neural networks for predicting bulk material properties.