MLatom

MLatom performs machine-learning–accelerated atomistic simulations and quantum chemical property modeling using kernel-based regression and molecular descriptors.

Key Features:

• Machine Learning Integration: Implements kernel ridge regression with support for Gaussian, Laplacian, and Matérn kernels.
• Input Representation: Accepts arbitrary input vectors and provides built-in molecular descriptors to convert molecular geometries into algorithm-ready inputs.
• Model Management: Saves and reuses trained ML models and estimates the generalization error of ML setups.
• Sampling and Gradient Calculation: Supports various sampling procedures and computes gradients of output properties.
• Implementation and Performance: Written primarily in Fortran, uses standard linear algebra libraries, and is optimized for shared-memory parallel computations.

Scientific Applications:

• Quantum Chemistry Simulations: Enables simulation of atomistic systems with improved computational efficiency using ML models.
• Molecular Dynamics: Facilitates detailed studies of molecular dynamics through ML-predicted forces and gradients.
• Electronic Structure Calculations: Supports workflows involving electronic structure data by providing ML-based surrogate models.
• Atomistic Modeling and Property Exploration: Assists exploration of molecular properties and behaviors at atomistic resolution.

Methodology:

Kernel ridge regression with Gaussian, Laplacian, and Matérn kernels; use of arbitrary input vectors or built-in molecular descriptors; sampling procedures and gradient evaluations; saving/reusing trained models and estimating generalization error; implementation in Fortran using standard linear algebra libraries and shared-memory parallelization.