FireGNN: Multi-Variant Graph Neural Models

• FireGNN is a collection of specialized graph neural network models tailored for distinct domains such as materials science, medical imaging, and wildfire prediction.
• It integrates innovations like relaxed equivariance, trainable fuzzy symbolic rules, feature interaction awareness, and multi-scale topological filtration for enhanced predictive accuracy and interpretability.
• Each variant has demonstrated state-of-the-art performance improvements over traditional methods, offering practical advantages in handling high-dimensional, sparse, or physically-constrained data.

FireGNN

FireGNN is a designation used for multiple independent graph neural network (GNN) architectures, each introduced for distinct domains and technical challenges across the literature. The principal FireGNN frameworks include: (1) the Force-Informed, Relaxed Equivariance GNN for rapid and accurate prediction of surface properties in materials science; (2) a neuro-symbolic GNN embedding trainable fuzzy rules for interpretable medical image classification; (3) a plug-and-play feature interaction-aware GNN tailored for high-dimensional sparse feature graphs; (4) a GCN-LSTM surrogate model for explainable global wildfire prediction; and (5) a filtration-based GNN for resilience optimization in pure-topology settings. Each approach is characterized by innovations that address symmetry, interpretability, higher-order structural priors, long-range dependencies, or feature sparsity, as detailed below.

1. FireGNN for Surface Property Prediction: Force-Informed, Relaxed Equivariance GNN

Problem Context and Motivation

The work function and cleavage energy are principal surface properties in materials modeling, relevant to electron emission, semiconductor devices, and heterogeneous catalysis. Density Functional Theory (DFT) yields accurate quantities but is computationally intractable for exhaustive high-throughput screening due to the combinatorial complexity of surfaces generated by composition, structure, orientation, and termination. A machine learning surrogate must retain the unique symmetry breaking of slabs—specifically, the reduction from 3D SO(3)/E(3) to 2D SO(2) due to surface-normal orientation—while capturing local geometric signals unencapsulated by invariant or fully equivariant models (Hsu et al., 22 Aug 2025).

Architectural Innovations

FIRE-GNN (Force-Informed, Relaxed Equivariance GNN) builds upon the SEGNN family, departing from standard approaches by:

• Explicitly concatenating the slab’s Cartesian $z$-coordinate as a node feature to model the broken symmetry along the surface normal, thereby reducing global symmetry to SO(2) and distinguishing top/bottom terminations.
• Incorporating node-level force information derived from a pretrained Orb-v3 universal machine learning interatomic potential (MLIP), supplying physically informed geometric cues to message passing.
• Training on a dataset of 33,631 slabs and 58,332 unique work function values, with DFT-computed targets for both work function and cleavage energy, enabling learning over a much larger and more chemically diverse domain than prior datasets.

Predictive and Generalization Performance

FIRE-GNN achieves a twofold reduction in mean absolute error for work function prediction (down to 0.065 eV for both top and bottom surfaces) over previous state-of-the-art models. Symmetry breaking in $z$ and force augmentation lead to consistent outperformance on both in-distribution and out-of-distribution splits (across parent structures, space groups, elements, and periodic table groups). For cleavage energy, symmetry breaking is less critical, but the integration of force features remains beneficial. The model exhibits strong out-of-distribution stability except in “periodic table group” splits, where the random forest baseline is comparable (Hsu et al., 22 Aug 2025).

Physical Rationale for Model Components

The injection of surface-normal information is required because the orientation of a surface directly impacts its electronic properties—forcing fully equivariant representations would confound top and bottom surfaces. MLIP-derived atomic forces serve as surrogates for local relaxation tendencies and bonding anisotropies near surfaces, enhancing sensitivity to orientation- and termination-specific features.

Broader Significance

FIRE-GNN enables DFT-accuracy-level surrogate screening of broad surface libraries, accelerating materials discovery for emission, catalysis, and interface design. The approach accommodates symmetry breaking naturally present in 2D materials and surfaces, and is extensible to usage as a generative design scoring function or rapid screening tool (Hsu et al., 22 Aug 2025).

2. FireGNN for Interpretable Medical Image Classification: Neuro-Symbolic GNN with Trainable Fuzzy Rules

Technical Motivation and Setting

Many GNN-based medical image classifiers lack intrinsic interpretability, with explanations arising only post hoc. FireGNN for medical imaging addresses this deficiency by embedding symbolic reasoning inside the model’s forward pass using fuzzy rules parameterized by learnable thresholds and sharpness, mapping topological descriptors to interpretable rule activations (Sengupta et al., 2 Sep 2025).

Core Framework and Reasoning Pipeline

The architecture constructs a graph where each node corresponds to a medical image and edges are derived by top-$k$ cosine similarity of high-level (ResNet18) image features. The main innovations are:

• Extraction of node-level topological descriptors: node degree, clustering coefficient, and 2-hop label agreement.
• Three trainable fuzzy rules, each activating via a sigmoid function with dataset-adaptive threshold and slope, producing a rule vector per node.
• Gating-based hybrid fusion of the fuzzy-rule-derived embedding and the standard neural embedding, enabling variable reliance on symbolic versus neural features per node.

Training and Evaluation

FireGNN is trained end-to-end with a cross-entropy objective, optionally augmented by auxiliary self-supervised losses for homophily and similarity entropy prediction. Benchmarks are provided on five MedMNIST datasets and the MorphoMNIST synthetic dataset, with strong accuracy improvements observed over base GNN models and auxiliary-task-enhanced variants. Representative gain: GCN base 88.2→ FireGNN 91.41 in OrganCMNIST accuracy (Sengupta et al., 2 Sep 2025).

Interpretability and Symbolic Adaptivity

Rule-based explanations are explicit—FireGNN produces logical statements such as IF degree $\geq7.28$ THEN high connectivity, with all rule parameters adapted to dataset statistics. This mode of operation is in contrast with post hoc explainers such as GNNExplainer, as the symbolic elements are intrinsic to prediction.

Current Limitations

The approach is presently restricted to three simple topological features. The rule set could be enriched (e.g., betweenness, eigenvector centrality) for greater coverage; computational cost and memory usage are also cited as limitations on large graphs (Sengupta et al., 2 Sep 2025).

3. FI-GNN: Feature Interaction-aware Graph Neural Networks

Addressed Challenge

Graphs arising in social networks, information retrieval, or recommendation often present high-dimensional, sparse node features (e.g., bag-of-words representations). Standard GNNs, relying on direct feature aggregation, yield weak node embeddings in this regime. FI-GNN introduces a plug-and-play module to explicitly encode and aggregate pairwise feature interactions, conditioning not only on the node but also on its context (Ding et al., 2019).

Architecture Overview

FI-GNN combines three modules:

• Message aggregator: standard GNN backbone (GCN or GraphSAGE).
• Feature factorizer: dense embeddings for each feature and explicit enumeration of all active pairwise interactions.
• Personalized attention: softmax-based weighing of interactions per node, aggregating the most informative interactions into the final node embedding.

The combined embedding is concatenated and input to downstream tasks such as node classification or link prediction.

Empirical Results

Demonstrated on BlogCatalog, Flickr, ACM, and DBLP, FI-GNN variants (FI-GCN, FI-GraphSAGE) consistently outperform their underlying backbones and factorization machine/non-GNN methods in both semi-supervised classification and link prediction. Gains are largest when attribute-label dependence is high, as in social graphs (Ding et al., 2019).

Relationship to Factorization Machines

The architecture generalizes factorization machines to the graph domain. When personalized attention and neighborhood dependencies are neglected, FI-GNN reduces to a standard factorization machine scoring function. This positions FI-GNN as a graph-aware enhancement to generalized factorization-based embedding approaches (Ding et al., 2019).

4. FireGNN for Explainable Global Wildfire Prediction (GCN-LSTM Surrogate)

Motivation and Problem Formulation

Conventional CNNs for wildfire modeling are hampered by local kernel receptive fields and spurious oceanic inputs (missing data). FireGNN addresses these by constructing a land-only graph with edges representing strong long-range correlations in wildfire time series, thereby bypassing the need for densely gridded 2D images (Chen et al., 2024).

Model Architecture

• Spatial component: single-layer GCN over a graph of $\sim$7,771 terrestrial nodes, with correlation-thresholded edges.
• Temporal component: LSTM models monthly sequences of GCN outputs, yielding multi-step forecasts of burnt area fraction.

Community structure is extracted post hoc via the Louvain modularity algorithm, providing geographically coherent clusters of wildfire correlation, and feature/node attributions are generated with Integrated Gradients to explain driver variables at various forecast horizons.

Quantitative Improvements and Explainability

FireGNN achieves superior metrics (MSE, RRMSE, SSIM, PSNR) compared to LSTM, ConvLSTM, and CAE-LSTM baselines when tested on unseen simulation ensembles. The method identifies known climate–wildfire teleconnections and regional fire regimes. Attributions reveal correct physical drivers, e.g., humidity as a negative predictor at short horizons, temperature positive at long horizons (Chen et al., 2024).

Limitations

The approach is validated on JULES-INFERNO simulations under LGM conditions, not observed fire data, and thus requires future work for real-world generalization and data assimilation (Chen et al., 2024).

5. FireGNN for Topology-Only RL-Based Network Resilience Optimization

Problem Setting and Deficiency of Standard GNNs

In edge-rewiring-based network resilience optimization tasks where inputs are featureless graphs (pure topology), standard and even advanced GNNs (GIN, DE-GNN, $k$-GNN) fail to yield distinguishable or informative state representations, hampering RL policy learning (Yang et al., 2021).

Filtration-Enhanced Graph Neural Network Design

FireGNN in this context is a filtration-based GNN inspired by persistent homology:

• Constructs a sequence of nested subgraphs by iteratively removing the highest-degree node, capturing graph evolution across multiple scales.
• Runs a GIN backbone on each subgraph; aggregate node, edge, and graph embeddings via attention-weighted linear combinations across the filtration sequence.
• Embedding aggregation enables the RL agent to differentiate actions and avoid loops in the solution trajectory, overcoming state aliasing intrinsic to pure-topology encoders.

Integration with RL and Performance

Embedded in the ResiNet RL framework for degree-preserving edge rewiring, FireGNN enables robust gains in both resilience and utility metrics versus transductive, non-learning (HC, Greedy, SA, EA) and learning-based (DE-GNN, $k$-GNN) baselines. Filtration order ablation confirms that the multi-scale subgraph representation is essential; using a standard GNN base yields only minimal gains (Yang et al., 2021).

Limitations

Accuracy of evaluation oracles and computational scalability are cited as bottlenecks for very large graphs; practical deployment is thus restricted to moderate scales unless future acceleration is achieved (Yang et al., 2021).

Summary Table: FireGNN Variants and Core Innovations

Variant	Domain	Central Innovation
FIRE-GNN (Hsu et al., 22 Aug 2025)	Surface property prediction	Relaxed equivariance; MLIP-derived forces
FireGNN (Sengupta et al., 2 Sep 2025)	Medical image classification	Intrinsic fuzzy symbolic rules
FI-GNN (Ding et al., 2019)	Sparse-feature graphs	Personalized interaction-aware embeddings
FireGNN (Chen et al., 2024)	Global wildfire prediction	Graph over land-only teleconnections; GCN-LSTM
FireGNN (Yang et al., 2021)	Network topology resilience	Filtration-based multi-scale encoding

Conclusion

The FireGNN nomenclature encompasses GNN architectures differentiated by domain, input structure, and inductive bias. Despite methodological diversity, each instantiates a nontrivial extension of standard GNNs: modeling physical symmetry breaking, integrating symbolic reasoning, learning personalized feature combinations, encoding spatial–temporal correlations, or capturing multi-scale topology in featureless graphs. The sum effect is to substantially advance the state of the art—either in predictive accuracy, interpretability, or generalizability—over baseline neural and heuristic methods in their respective application areas.