Atom Centered Symmetry Functions Selection for Neural Network Potentials
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for developing High Dimensional Neural Network Potentials (HDNNPs) for chemical systems lack a systematic approach to identify optimal Atom Centered Symmetry Functions (ACSFs), relying on trial-and-error heuristics, which hinders the development of accurate and transferable models.
Innovation Solution
A processor-implemented method and system that generates, prunes, and sorts ACSFs based on their distribution and pairwise distance to determine an optimal set, ensuring rotational, translational, and permutational invariance, and trains HDNNPs using the identified optimal ACSFs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If trial-and-error approach based on heuristics is used to identify ACSFs, then development process is simple to implement, but accuracy and transferability of HDNNPs cannot be ensured
Solution Approach 1:
The patent systematically varies ACSF parameters (eta, rs, zeta, lambda, Rc) across multiple orders of magnitude to generate a comprehensive set of candidate functions. This parameter exploration transforms the trial-and-error approach into a systematic parameter sweep, enabling identification of optimal parameter combinations that maximize HDNNP accuracy while maintaining rotational, translational, and permutational invariance.
Solution Approach 2:
The patent divides the ACSF identification process into distinct stages: (1) generating initial candidate ACSFs by varying parameters, (2) computing ACSF values for each local environment in training data, (3) constructing histograms to analyze distribution characteristics, (4) pruning based on width and maximum value criteria, (5) sorting by spread metric, and (6) selecting optimal set. This segmentation transforms the monolithic trial-and-error process into manageable systematic steps.
2Manufacturing precision
If comprehensive parameter sweep is performed to identify optimal ACSFs, then accuracy of HDNNP is improved, but computational time and complexity increase
Solution Approach 1:
The patent performs preliminary filtering of ACSF candidates by computing their values across the training dataset and analyzing distribution characteristics (histograms) before final model training. By pre-evaluating and pruning ACSFs based on width and maximum value criteria, the method eliminates unpromising candidates early, reducing the computational burden of subsequent training iterations while preserving accuracy-critical functions.
Solution Approach 2:
The patent extracts and removes ACSF candidates that do not meet predefined criteria (narrow distributions, excessive maximum values) from the candidate set. This extraction of undesirable elements prunes the search space systematically, retaining only ACSFs with appropriate distribution characteristics that are likely to contribute meaningfully to HDNNP accuracy, thereby reducing computational overhead.
3Manufacturing precision
If large number of ACSFs are used to capture complex chemical environments, then model accuracy improves, but model complexity and training difficulty increase
Solution Approach 1:
The patent employs feedback mechanisms by computing the spread metric for each ACSF based on its distribution across the training dataset and using this information to guide selection. The histogram analysis provides feedback on ACSF effectiveness, and the pruning criteria (width and maximum value thresholds) use this feedback to iteratively refine the candidate set, retaining only ACSFs that provide meaningful discriminatory power for different chemical environments.
Solution Approach 2:
The patent generates an excessive number of initial ACSF candidates by sweeping parameters across wide ranges, then systematically prunes this oversized set to the optimal subset. This approach of generating more candidates than ultimately needed ensures comprehensive coverage of the ACSF parameter space, allowing the pruning process to select the most effective functions while discarding redundant ones, achieving optimal model complexity.
4Manufacturing precision
If systematic methodology is developed for identifying ACSFs, then accuracy and transferability are improved, but method complexity and implementation difficulty increase
Solution Approach 1:
The patent develops a universal framework for ACSF identification that can be applied across different chemical systems and HDNNP applications. The systematic parameter sweep methodology, histogram-based analysis, and pruning criteria constitute a transferable protocol that works for diverse chemical environments, making the method universally applicable while maintaining systematic rigor. The approach unifies ACSF selection under consistent principles regardless of the specific chemical system.
Data Source
AI summary
This disclosure relates generally to method to determine an optimal set of atom centered symmetry functions. One or more parameters associated with one or more atom centered symmetry functions (ACSFs) are received. An initial set of ACSFs is generated by varying the one or more parameters. A histogram with a prespecified bin size is constructed to obtain a distribution of value of each of the initial set of ACSFs. A pruned list of ACSFs is obtained based on width and maximum value of the distribution of the value of initial set of ACSFs. The pruned list of ACSFs is sorted in decreasing order of spread to obtain a sorted list of ACSFs. An optimal set of one or more shortlisted ACSFs is determined by traversing through the sorted list of ACSFs. A high dimensional neural network potential is trained based on the optimal set of one or more shortlisted ACSFs.