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

2020/12/31 by Suresh Kondati Natarajan, Natarajan, Suresh Kondati, Caro, Miguel A. +1 · 1 citation
Chemistry · Computer Science · Materials Science · #Computational Drug Discovery Methods #Crystallography and molecular interactions #FOS: Computer and information sciences #FOS: Physical sciences #Machine Learning (cs.LG) #Machine Learning in Materials Science #Materials Science (cond-mat.mtrl-sci)

paper · pdf · doi:10.48550/arxiv.2101.00049

openalex publication_date 2020/12/31 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

Modeling non-empirical and highly flexible interatomic potential energy\nsurfaces (PES) using machine learning (ML) approaches is becoming popular in\nmolecular and materials research. Training an ML-PES is typically performed in\ntwo stages: feature extraction and structure-property relationship modeling.\nThe feature extraction stage transforms atomic positions into a\nsymmetry-invariant mathematical representation. This representation can be\nfine-tuned by adjusting on a set of so-called "hyper-parameters" (HPs).\nSubsequently, an ML algorithm such as neural networks or Gaussian process\nregression (GPR) is used to model the structure-PES relationship based on\nanother set of HPs. Choosing optimal values for the two sets of HPs is critical\nto ensure the high quality of the resulting ML-PES model.\n In this paper, we explore HP optimization strategies tailored for ML-PES\ngeneration using a custom-coded parallel particle swarm optimizer (available\nfreely at https://github.com/suresh0807/PPSO.git). We employ the smooth overlap\nof atomic positions (SOAP) descriptor in combination with GPR-based Gaussian\napproximation potentials (GAP) and optimize HPs for four distinct systems: a\ntoy C dimer, amorphous carbon, \α-Fe, and small organic molecules (QM9\ndataset). We propose a two-step optimization strategy in which the HPs related\nto the feature extraction stage are optimized first, followed by the\noptimization of the HPs in the training stage. This strategy is computationally\nmore efficient than optimizing all HPs at the same time by means of\nsignificantly reducing the number of ML models needed to be trained to obtain\nthe optimal HPs. This approach can be trivially extended to other combinations\nof descriptor and ML algorithm and brings us another step closer to fully\nautomated ML-PES generation.\n

Citations

• Gaussian approximation potentials: A brief tutorial introduction
• Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning
• Self-Consistent Equations Including Exchange and Correlation Effects
• Quantum-chemical insights from deep tensor neural networks
• DScribe: Library of descriptors for machine learning in materials science

Cited by

• Machine learning model for efficient nonthermal tuning of the charge density wave in monolayer NbSe₂

Related

• Data-efficient multi-fidelity training for high-fidelity machine\n learning interatomic potentials
• Fine-Tuning Bulk-oriented Universal Interatomic Potentials for Surfaces: Accuracy, Efficiency, and Forgetting Control
• Global Universal Scaling and Ultra-Small Parameterization in Machine Learning Interatomic Potentials with Super-Linearity
• Augmenting Molecular Graphs with Geometries via Machine Learning Interatomic Potentials
• Performance and Cost Assessment of Machine Learning Interatomic Potentials