Abstract:Descriptors are physically-inspired schemes for representing atomistic systems that play a central role in the construction of models of potential energy surfaces. Although physical intuition can be flexibly encoded into descriptor schemes, they are generally ultimately guided only by the spatial or topological arrangement of atoms in the system. Here, we propose a novel approach for the optimization of descriptors based on encoding information about geodesic distances along potential energy manifolds into the hyperparameters of commonly used descriptor schemes. To accomplish this, we combine two ideas: (1) a differential-geometric approach for the fast estimation of approximate geodesic distances; and (2) an information-theoretic evaluation metric - information imbalance - for measuring the shared information between two distance measures. Using the MD22 datasets of ethanol, malonaldehyde, and aspirin, we first show that Euclidean (in Cartesian coordinates) and geodesic distances are inequivalent distance measures, indicating the need for updated ground-truth distance measures that go beyond the Euclidean distance. We then utilize a Bayesian optimization framework to show that descriptors (in this case, atom-centered symmetry functions) can be optimized to maximally express a certain type of distance information, such as Euclidean or geodesic information. We also show that modifying the Bayesian optimization algorithm to minimize a combined Euclidean+geodesic objective function can yield descriptors that not only express both Euclidean and geodesic distance information simultaneously, but in fact resolve substantial disagreements between descriptors optimized to encode only one type of distance measure. We discuss the relevance of our approach to the design of more physically rich and informative descriptors that can encode useful, alternative information about molecular systems.

Optimizing many-body atomic descriptors for enhanced computational performance of machine learning based interatomic potentials

Learning Effective Molecular Models from Experimental Observables.

Fast proper orthogonal descriptors for many-body interatomic potentials

Improve the Performance of Machine-Learning Potentials by Optimizing Descriptors

Comparing molecules and solids across structural and alchemical space

Optimal design of experiments in the context of machine-learning inter-atomic potentials: improving the efficiency and transferability of kernel based methods

Investigating 3D Atomic Environments for Enhanced QSAR

Atomistic Descriptor Optimization Using Complementary Euclidean and Geodesic Distance Information

Particle Swarm Based Hyper-Parameter Optimization for Machine Learned Interatomic Potentials

Extending the atomic decomposition and many-body representation, a chemistry-motivated monomer-centered approach for machine learning potentials

Feature Optimization for Atomistic Machine Learning Yields A Data-Driven Construction of the Periodic Table of the Elements

Continuous and optimally complete description of chemical environments using Spherical Bessel descriptors

Expanding Density-Correlation Machine Learning Representations for Anisotropic Coarse-Grained Particles

Atomic Structure Optimization with Machine-Learning Enabled Interpolation between Chemical Elements

Driving atomic structures of molecules, crystals, and complex systems with local similarity kernels

Designing hybrid descriptors for improved machine learning models in atomistic materials science simulations

Efficient and Robust Parameter Optimization of the Unitary Coupled-Cluster Ansatz

Bayesian optimization of atomic structures with prior probabilities from universal interatomic potentials

Machine-learning accelerated geometry optimization in molecular simulation

Simultaneous Optimization of Biomolecular Energy Functions on Features from Small Molecules and Macromolecules.

Efficient and Precise Force Field Optimization for Biomolecules Using DPA-2