Probing Accuracy-Speedup Tradeoff in Machine Learning Surrogates for Molecular Dynamics Simulations

Area of Science:

• Computational Materials Science
• Soft Matter Physics
• Machine Learning Applications

Background:

• Machine learning (ML) surrogates offer performance gains for molecular dynamics (MD) simulations of soft materials.
• Acquiring large training datasets is often necessary for high-accuracy ML surrogates, posing a challenge with limited high-performance computing (HPC) resources.
• Optimizing training dataset size is crucial for efficient ML surrogate development.

Purpose of the Study:

• To investigate the accuracy-computational speed tradeoff in ML surrogates for MD simulations of confined electrolytes.
• To determine the optimal training dataset size under HPC constraints.
• To evaluate the generalizability of ML surrogates and develop a metric for net speedup.

Main Methods:

• An artificial neural network (ANN) based surrogate model was developed for MD simulations.
• Accuracy was quantified using root-mean-square errors (RMSE) against MD ground truth.
• Computational performance was assessed via speedup, including training data acquisition time.
• Surrogate generalizability was tested on unseen data, and a net speedup metric was introduced.

Main Results:

• Increased surrogate accuracy correlated with a decrease in computational speedup.
• Speedup loss was inversely proportional to the training dataset size.
• The study identified a direct tradeoff between model accuracy and computational efficiency.
• Generalizability analysis revealed incurred errors on interpolated inputs.

Conclusions:

• A balance between accuracy and computational gains is essential when developing ML surrogates for MD simulations.
• Training dataset size critically influences this accuracy-speedup tradeoff.
• The developed net speedup metric aids in evaluating the practical computational benefits of ML surrogates.