Introducing Neural Mesh for Logical Synthesis Models

Abstract

The paper introduces the Neural Mesh, a novel architecture that serves as an alternative to conventional neural network models widely used in contemporary AI systems. In parallel, the paper introduces the term Logical Synthesis Model (LSM) as a complement to Large Language Models (LLMs). While LLMs primarily rely on statistical representations and language-based reasoning to acquire and generate knowledge, Neural Mesh-based LSMs focus on structured representations and engineering reasoning to represent, analyze, synthesize, and control systems. LLMs excel atlearning and reasoning over human knowledge expressed through language, whereas LSMs aim to learn structured representations that support the “understanding” of physical systems. In this sense, LLMs and LSMs may play roles analogous to the cerebrum and the cerebellum in biological intelligence. The novelty of the paper lies in bridging and combining the concepts of TP models, TP model transformation, and neural-network architectures. The proposed Neural Mesh is mathematically equivalent to the well-established TP model function family. The TP model transfor-mation provides a framework for constructing TP models with unique features that are particularly advantageous for systems and control applications. The contribution of the Neural Mesh is that it represents these unique features in the form of a special neural-network architecture, where these features are expressed as trainable neural connections and parameters. As a result, elements that are traditionally determined through an offline TP model transformation become directly tunable through learn-ing, while preserving the mathematical structure and control-theoretic advantages of the TP model framework. Therefore, the Neural Mesh representation opens a future research direction in which the TP model transformation is effectively embedded into the training process itself. Rather than generating a TP model through a subsequent offline transformation, the corresponding TP-model components are directly constructed and tuned during system identification via neural-network training tools.