Although molten salt reactors and thermal storage systems are attracting increasing interest, our understanding of the physicochemical properties of molten salts is still incomplete. This is largely due to the difficulty of conducting experiments under extreme temperatures with strict control of impurities and corrosion. Ab initio calculations, machine-learned force fields, and classical molecular dynamics have helped to alleviate some of these issues. However, discrepancies between experimental and theoretical computations of the thermal conductivity of fluoride molten salts have become of increasing concern. In this paper, we present a modernized method for training a transferable equivariant graph neural network force fields to model a simple fluoride molten salt system, LiF–NaF, using minimal ab initio calculations. Using this transferable machine-learned force field, the thermal conductivity as well as various other functions of LiF–NaF were computed at various chemical temperatures and ratios in order to gain new insights into the limitations and behaviors of molten salts in relation to their thermal conductivity. Results show discrepancies between experimental and theoretical computations of the thermal conductivity as a function of temperature but good agreement between experimental and theoretical computations of the thermal conductivity as a function of ratio. Secondary results show compelling agreement of a machine-learned force field with first-principles computations and the ability to interpolate and extrapolate various chemical ratios.
A bstract: Graph neural networks have recently shown strong promise for event reconstruction tasks in Liquid Argon Time Projection Chambers, yet their performance remains limited for underrepresented classes of particles, such as Michel electrons. In this work, we investigate physics-informed strategies to improve semantic segmentation within the NuGraph2 architecture. We explore three complementary approaches: (i) enriching the input representation with context-aware features derived from detector geometry and track continuity, (ii) introducing auxiliary decoders to capture class-level correlations, and (iii) incorporating energy-based regularization terms motivated by Michel electron energy distributions. Experiments on MicroBooNE public datasets show that physics-inspired feature augmentation yields the largest gains, particularly boosting Michel electron precision and recall by disentangling overlapping latent space regions. In contrast, auxiliary decoders and energy-regularization terms provided limited improvements, partly due to the hit-level nature of NuGraph2, which lacks explicit particle- or event-level representations. Our findings highlight that embedding physics context directly into node-level inputs is more effective than imposing task-specific auxiliary losses, and suggest that future hierarchical architectures such as NuGraph3, with explicit particle- and event-level reasoning, will provide a more natural setting for advanced decoders and physics-based regularization. The code for this work is publicly available on Github at https://github.com/vitorgrizzi/nugraph_phys/tree/main_phys .
28 May 2026 in Machine Learning: Science and Technology
With the growing popularity of artificial intelligence (AI) used for scientific applications, the ability of attribute a result to a reasoning process from the network is in high demand for robust scientific generalizations to hold. In this work we aim to motivate the need for and demonstrate the use of post-hoc explainability methods when applied to AI methods used in scientific applications. To this end, we introduce explainability add-ons to the existing graph neural network (GNN) for neutrino tagging, NuGraph2. The explanations take the form of a suite of techniques examining the output of the network (node classifications) and the edge connections between them, and probing of the latent space using novel general-purpose tools applied to this network. We show how none of these methods are singularly sufficient to show network ‘understanding’, but together can give insights into the processes used in classification. While these methods are tested on the NuGraph2 application, they can be applied to a broad range of networks, not limited to GNNs. The code for this work is publicly available on GitHub at https://github.com/voetberg/XNuGraph .
11 August 2026 in Journal of Chemical Theory and Computation