Improved loss functions for machine-learned atomic potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning Applications

Background:

• Machine learning (ML) is increasingly vital in scientific research, particularly in chemistry for predicting molecular and material properties.
• Training ML models, such as neural network potentials (NNPs), requires careful consideration of dataset quality and loss functions.

Purpose of the Study:

• To investigate the impact of different loss functions on the training and performance of neural network potentials (NNPs).
• To develop a novel loss function that enhances the accuracy and generality of NNPs.

Main Methods:

• Tested mean-squared error and Huber loss functions for training NNPs.
• Derived and evaluated a new Asinh-based loss function.
• Analyzed the effect of loss functions on error minimization and parameter gradients during NNP training.

Main Results:

• Both Huber and Asinh loss functions demonstrated improved NNP training by minimizing errors and anomalies.
• The Asinh loss function yielded significant gains in NNP accuracy and generality.
• Optimized training led to NNPs with greater effective dimensionality.

Conclusions:

• The choice of loss function critically impacts NNP training and predictive capabilities.
• The novel Asinh loss function offers a superior approach for developing accurate and generalizable NNPs in chemistry.
• Minimizing errors during optimization is key to enhancing NNP performance.