Smooth Overlap of Atomic Positions (SOAP)

SOAP represents each neighboring atom not as a discrete point, but as a Gaussian-smeared density centered at its position. This smoothing ensures the descriptor is continuously differentiable with respect to atomic positions, a critical requirement for computing forces in molecular dynamics. The width of the Gaussian is controlled by a hyperparameter that defines the effective radius of atomic influence.

The core mathematical trick of SOAP is expanding the local density field in a basis of spherical harmonics and radial basis functions, then computing the power spectrum of the resulting expansion coefficients. This operation guarantees rotational invariance—the descriptor remains identical regardless of how the molecule is oriented in 3D space. The power spectrum is formed by contracting the expansion coefficients over the angular momentum index.

Because SOAP constructs a total neighbor density by summing over all atoms within a cutoff radius, the descriptor is inherently permutationally invariant. Swapping the labels of two identical atoms in the input list produces exactly the same SOAP vector. This property reflects the physical indistinguishability of identical atoms and eliminates the need for data augmentation during training.

SOAP descriptors are often compared using a dot-product kernel that quantifies the similarity between two atomic environments. This kernel can be raised to a power to sharpen the similarity measure, effectively controlling how local or global the comparison is. The resulting SOAP kernel is rotationally invariant, permutationally invariant, and smoothly differentiable, making it suitable for Gaussian Process Regression models like GAP.

The accuracy of the SOAP representation is controlled by two truncation parameters: the maximum radial quantum number and the maximum angular momentum quantum number. Increasing these values systematically expands the basis set, allowing the descriptor to capture finer chemical details. In the limit of a complete basis, the SOAP power spectrum uniquely determines the local atomic environment up to rotation.

For systems containing multiple chemical elements, SOAP is extended by constructing partial densities for each species. The power spectrum is then computed for each pair of species, yielding a descriptor that distinguishes not only the geometric arrangement but also the chemical identity of neighbors. This multi-species SOAP vector grows quadratically with the number of elements but provides a complete description of the local chemical environment.

A systematic and complete framework for constructing permutationally and rotationally invariant atomic descriptors. ACE provides the theoretical foundation for a family of highly efficient machine learning interatomic potentials.

• Systematic: Can be systematically improved by increasing the body-order of the expansion
• Complete: Capable of representing any smooth function of atomic coordinates to arbitrary accuracy
• Relationship to SOAP: SOAP can be understood as a specific truncation of the ACE formalism, using a spherical harmonic expansion of the neighbor density

A fundamental design constraint for machine learning models of atomic systems. The prediction must remain unchanged when the order of identical atoms in the input list is swapped.

• Physical basis: Atoms of the same element are indistinguishable quantum mechanically
• SOAP's solution: The power spectrum and bispectrum are constructed to be invariant under permutations of identical atoms
• Contrast: A naive Cartesian coordinate input would change if atom indices are swapped, leading to unphysical predictions

A neural network architecture that guarantees its output transforms predictably under symmetry operations of 3D space, such as rotation and translation.

• Invariance vs. Equivariance: SOAP descriptors are invariant (unchanged by rotation), while equivariant networks can predict vectorial properties like forces that rotate with the molecule
• Complementary approaches: SOAP is often used as input to invariant models for energy prediction, while equivariant networks learn these symmetries internally
• Modern trend: Equivariant architectures like NequIP and MACE are increasingly preferred, but SOAP remains a benchmark descriptor

A training paradigm where the loss function directly compares the atomic forces predicted by the model to reference forces from quantum mechanical calculations.

• Training signal: Forces provide 3N times more training data per configuration than a single energy value
• SOAP integration: SOAP descriptors are used to featurize each atom's environment, and the model predicts forces as the negative gradient of the predicted energy with respect to atomic positions
• Physical consistency: Force matching ensures the model learns a conservative potential energy surface where forces are the true derivatives of the energy

A learning strategy where a model is trained to predict the small difference between a low-level, inexpensive theory and a high-level, accurate theory.

• Efficiency: Combines the speed of a baseline method like DFTB with the accuracy of coupled cluster
• SOAP's role: SOAP descriptors capture the local atomic environment, and the model learns a correction term ΔE = E_high - E_low
• Transferability: The correction learned with SOAP descriptors often transfers well across similar chemical systems, reducing the need for expensive reference calculations

The process of assigning a confidence interval to a machine learning model's prediction, critical for assessing reliability and guiding data acquisition.

• Gaussian Process Regression: SOAP is frequently paired with Gaussian processes, which provide a natural variance estimate alongside the prediction
• Active learning: When a SOAP-based model encounters an atomic environment with high uncertainty, it can flag that configuration for a new quantum mechanical calculation
• Extrapolation detection: The smooth overlap kernel naturally identifies when a new environment lies far from the training distribution, preventing overconfident extrapolation