graph neural networks

Intelligent scheduling of aerospace precision forgings based on deep reinforcement learning and graph neural networks

Data-efficient exploration of atomic clusters via uncertainty quantification on complex potential energy surfaces

Machine learning interatomic potentials have become an effective method for exploring complex potential energy surfaces; however, their application to atomic clusters is frequently hindered by the high cost of sampling diverse isomer spaces and the difficulty in ensuring model generalizability across complex energy landscapes. While uncertainty quantification (UQ) offers a pathway to mitigate data scarcity, its efficacy in capturing continuous potential energy surface features and guiding active learning within the complex landscape of clusters remains systematically unverified. In this study, we developed and evaluated three different UQ frameworks based on advanced UQ methods and integrated with graph neural networks: Bayesian Neural Networks, Evidence Neural Networks (ENN), and Monte Carlo Dropout (MCD). We first validated these models on the MD17 dataset to establish baseline performance, followed by a rigorous assessment on complex cluster systems (Ta2N3- and LaSi24) to probe their decision-making mechanisms in high-dimensional spaces. Our results show that ENN effectively reflects data adequacy, while MCD exhibits excellent robustness, with an MSE of <0.3 eV2 in the first 20 epochs. Furthermore, active learning driven by MCD significantly reduces the computational overhead of first-principles calculations while maintaining high predictive accuracy. This study provides a physics-informed guideline for selecting UQ strategies, facilitating the autonomous and efficient discovery of stable cluster isomers.

Neural Network-Based Accelerated Model for Condensed Combustion Products in Solid Propellant

Graph Neural Network and Artificial Based Models in Preclinical Pharmacology: Predictivity, Opportunities and Challenges

Major challenges in preclinical pharmacology, such as low predictivity, high rates of attrition, prolonged development cycles, and high costs remain significant issues as they largely relate to how translational relevance of past experimental models. This work aims to assess the potential of artificial intelligence-based models, especially graph neural networks in enhancing the predictivity, opportunity, discovery and challenges solving in preclinical pharmacology. This is a synthesis review on machine learning, deep learning and graph-based architectures being used to predict molecular properties, toxicity, drug-target and drug-drug interaction, drug repurposing and de-novo molecule design and their methodological basis and application in preclinical workflows. The reviewed literature reveals that AI models, particularly graph neural networks, may be useful in learning complex chemical and biological relationships, leading to improved prediction of ADMET properties, toxicity endpoints, pharmacokinetics, and therapeutic efficacy, while reducing reliance on animal models and accelerating candidate prioritization. Despite these advances, significant limitations persist, including data scarcity and heterogeneity, limited interpretability, bias, scalability constraints, and challenges in generalization across biological systems and regulatory acceptance. In conclusion, artificial intelligence and graph neural network–based approaches represent a transformative paradigm for preclinical pharmacology by enhancing predictivity, efficiency, and translational relevance; however, their successful integration into drug development will depend on advances in high-quality data generation, explainable and hybrid modeling strategies, standardized validation practices, and ethical and regulatory alignment

Multi-Modal Knowledge Graph Reasoning Engine

This paper introduces a novel Multi-Modal Knowledge Graph Reasoning Engine designed to facilitate complex reasoning and knowledge discovery through the interactive integration of diverse knowledge graph modalities. The core claim is that leveraging the interplay between multi-modal knowledge graphs significantly enhances the ability to perform sophisticated reasoning tasks compared to traditional, single-modality approaches. The proposed mechanism centers around constructing a fused knowledge graph encompassing textual, visual, and auditory data, coupled with a graph neural network (GNN)-based reasoning engine. This architecture overcomes the limitations inherent in single-modality knowledge graphs, offering a more comprehensive and nuanced understanding of interconnected information. The engine is designed to dynamically adapt to the relationships between different modalities, improving the accuracy and efficiency of reasoning processes. The paper outlines the architecture, the GNN implementation, and potential applications, illustrating a pathway toward intelligent systems capable of sophisticated knowledge extraction and inference.

基于神经符号人工智能的知识推理

This paper explores a novel approach to artificial intelligence by integrating the strengths of neural networks and symbolic reasoning. The core idea is to leverage the pattern recognition capabilities of neural networks for knowledge extraction and representation, coupled with the logical deduction and inference capabilities of symbolic reasoning systems. This hybrid architecture aims to construct AI systems that not only possess high accuracy but also offer enhanced interpretability and reliability. The system utilizes a neural network to process raw input data and generate a structured knowledge graph, which is then subjected to symbolic reasoning algorithms to derive conclusions and make decisions. The proposed method addresses the limitations of existing neural AI approaches that often lack explicit knowledge representation and reasoning mechanisms, leading to a lack of transparency and difficulty in debugging. This research contributes to the development of more robust and trustworthy AI systems by grounding AI decisions in a formally represented knowledge base. The evaluation framework will focus on assessing the system's accuracy, interpretability, and robustness across various reasoning tasks.

Dynamic Topology Neural Networks (DTNNs) Learning

This paper introduces Dynamic Topology Neural Networks (DTNNs), a novel approach to neural network design that leverages continuous topological evolution and feedback learning to autonomously construct and optimize network topologies. Unlike traditional static neural networks, DTNNs adapt their structure in real-time, responding to evolving input data and task requirements. The core mechanism involves a graph-based neural network framework where neurons and connections form a dynamic topology. Each neuron possesses a state vector, influencing its activation level, connection weights, and local environmental information. Topological evolution is governed by mechanisms, such as genetic algorithms or reinforcement learning, dynamically adjusting the network's structure by adding, removing, or modifying connections based on neuron states and network performance metrics. Simultaneously, feedback learning updates neuron activation levels through gradient descent or pulse feedback, driven by input signals and performance objectives. This architecture offers enhanced adaptability, robustness, and scalability, surpassing the limitations of static networks in handling unstructured and dynamic data. The key innovation lies in the dynamic control of topology and the deep learning of neuron states, providing a more flexible and powerful framework for complex data processing.

Temporal Graph Neural Networks with Adaptive Relational Strength

This paper introduces a novel approach to Graph Neural Networks (GNNs) designed to effectively capture temporal dependencies within dynamic graph structures. Traditional GNNs often fail to adequately represent the evolving nature of relationships between nodes, hindering their performance in systems where the graph topology changes over time. Our proposed Temporal Graph Neural Networks (T-GNNs) address this limitation by incorporating a 'temporal strength' factor into node embeddings. This factor is dynamically adjusted based on the observed evolution of relationships within the graph, learned through a recurrent attention mechanism that specifically focuses on relational changes. The core innovation lies in the adaptive adjustment of relational strength, allowing the network to prioritize relevant temporal information. We demonstrate through a theoretical analysis and structural design that this approach significantly improves the ability of GNNs to model time-evolving graph data. The resulting T-GNN architecture offers a robust framework for analyzing and predicting behavior in dynamic systems represented as graphs.

RGDLAN and HMARN: A dual-path graph neural approach with multi-agent systems for listed enterprise violation prediction

Title: Non-Local Graph Neural Networks for Semantic Representation

This paper explores the application of non-local graph neural networks to enhance semantic representation learning. Traditional graph neural networks primarily focus on node features, often struggling to capture the nuanced relationships within a graph. We propose a novel architecture that explicitly models the *meaning* of the graph, achieved through a mechanism that learns representations sensitive to the graph's structure and content. This approach aims to improve generalization and robustness compared to existing methods, particularly in scenarios with complex, high-dimensional graphs. The core of this work involves a modified attention mechanism integrated within the graph neural network to facilitate the propagation of information across the graph, thereby strengthening the representation of semantic connections. We present a series of experiments evaluating the performance of this architecture on various graph datasets, demonstrating significant improvements in semantic understanding and overall accuracy.

Temporal Graph Neural Networks with Adaptive Relational Strength

This paper introduces a novel approach to Graph Neural Networks (GNNs) designed to effectively capture temporal dependencies within dynamic graph structures. Traditional GNNs often fail to adequately represent the evolving nature of relationships between nodes, hindering their performance in systems where the graph topology changes over time. Our proposed Temporal Graph Neural Networks (T-GNNs) address this limitation by incorporating a 'temporal strength' factor into node embeddings. This factor is dynamically adjusted based on the observed evolution of relationships within the graph, learned through a recurrent attention mechanism that specifically focuses on relational changes. The core innovation lies in the adaptive adjustment of relational strength, allowing the network to prioritize relevant temporal information. We demonstrate through a theoretical analysis and structural design that this approach significantly improves the ability of GNNs to model time-evolving graph data. The resulting T-GNN architecture offers a robust framework for analyzing and predicting behavior in dynamic systems represented as graphs.

Modularity-Based Linkage Model Assisted Gradient Descent for Neural Network Training