NuGraph2 with Context-Aware Inputs: Physics-Inspired Improvements in Semantic Segmentation

TL;DR

NuGraph2's performance on LArTPC semantic segmentation is limited for underrepresented Michel electrons. The authors test physics-informed strategies: augmenting node inputs with context-aware features, adding auxiliary decoders for class-level correlations, and introducing energy-based regularization aligned with Michel energy distributions. The results show that context-aware input features furnish the largest improvements, particularly increasing Michel electron precision and recall by better disentangling latent regions, while auxiliary decoders and energy regularization yield little or negative gains due to the hit-level nature of NuGraph2. The study suggests that embedding physics context at the representation level is more effective in this architecture and that future NuGraph3 with explicit particle- and event-level reasoning could better leverage such inductive biases.

Abstract

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.

Figures (5)

• Figure 1: Node connectivity in the $time \ vs \ wire$ plane for the three different planes $u$, $v$, and $y$. Each node represents a detector hit and edges were constructed using Delaunay triangulation algorithm. Nodes are color-coded according to their true semantic label. The graph exhibits linear tracks characteristic of HIPs and MIPs, and dispersed nodes with high connectivity acting as "hubs". Both patterns served as inspiration for augmenting the input features.
• Figure 2: a) Precision and b) Recall confusion matrices for the baseline network.
• Figure 3: a) Precision and b) Recall confusion matrices for the network with physics-based extended features.
• Figure 4: Distribution of Michel electrons in the $Integral$$vs$ Deposited Energy summed across the three planes (event-wise). The reported $Integral$ is the sum of the integral of all hits with Michel as their true label in each event across all LArTPC planes, while Deposited Energy is the sum of the corresponding simulated energy deposits.
• Figure 5: Michel electron deposited energy distribution summed across all planes (i.e., for the entire event). The red curve indicates the Langauss fit, starting at $E=8.5$ MeV. See Fig. \ref{['fig:kde']} for a description of how the Deposited Energy is computed.