Uncertainty and Exploration of Deep Learning-based Atomistic Models for Screening Molten Salt Properties and Compositions

Stephen T. Lam,Shubhojit Banerjee,Rajni Chahal

2024-05-01

Abstract:Due to extreme chemical, thermal, and radiation environments, existing molten salt property databases lack the necessary experimental thermal properties of reactor-relevant salt compositions. Meanwhile, simulating these properties directly is typically either computationally expensive or inaccurate. In recent years, deep learning (DL)-based atomistic simulations have emerged as a method for achieving both efficiency and accuracy. However, there remain significant challenges in assessing model reliability in DL models when simulating properties and screening new systems. In this work, structurally complex LiF-NaF-ZrF$_4$ salt is studied. We show that neural network (NN) uncertainty can be quantified using ensemble learning to provide a 95% confidence interval (CI) for NN-based predictions. We show that DL models can successfully extrapolate to new compositions, temperatures, and timescales, but fail for significant changes in density, which is captured by ensemble-based uncertainty predictions. This enables improved confidence in utilizing simulated data for realistic reactor conditions, and guidelines for training deployable DL models.

Materials Science,Chemical Physics

What problem does this paper attempt to address?

This paper primarily explores the problem of handling uncertainty and exploring new scenarios in deep learning-based atomic models, particularly in the screening of molten salt properties and compositions. Due to extreme chemical, thermal, and radiation environments, existing databases of molten salt properties lack the necessary experimental thermodynamic data. Simulating these properties directly is both expensive and potentially inaccurate. In recent years, deep learning-based atomic simulation has provided an efficient and accurate approach. However, evaluating the reliability of the model under new conditions is a major challenge.

In the paper, the researchers use the LiF-NaF-ZrF4 molten salt with complex structure as an example to demonstrate the quantification of neural network (NN) uncertainty through ensemble learning, thereby providing predictions with a 95% confidence interval (CI). They found that the DL model can successfully extrapolate to new compositions, temperatures, and time scales, but fails in cases of significant density variation, which can be captured by uncertainty predictions based on ensembles. This approach enhances confidence in using simulated data under actual reactor conditions and provides guidelines for training deployable DL models.

The researchers employed density functional theory (DFT) calculations to train the deep learning potential and verified the accuracy of the model through comparison with DFT calculations. They also estimated uncertainty by constructing an ensemble of neural networks and obtained a conservative error estimate by minimizing the squared error of the top 5% error configurations.

Results show that the DL model performs well in terms of composition and temperature variations but may exhibit high errors when subjected to pressure and density changes caused by volume expansion. By quantifying uncertainty, the researchers are able to determine when the model's predicted uncertainty is low, allowing for reliable use of the model and reducing reliance on expensive DFT calculations. To reduce overall uncertainty, it is recommended to increase volumetric sampling when developing DL atomic models.

In summary, the paper addresses how to quantify and manage uncertainty in DL-driven atomic models to more accurately predict molten salt properties, particularly in the design and optimization of molten salt nuclear reactors.