Multimodal Intelligent Assessment and Dynamic Regulation of Financial and Tax Resilience in Manufacturing Enterprises under Supply Chain Shocks

Abstract

Supply chain disruptions and cyberattacks can jointly amplify production, liquidity, and tax risks in manufacturing enterprises, whereas conventional static scoring methods struggle to translate heterogeneous evidence into security responses. This paper proposes an artificial intelligence (AI)-driven framework that couples a sparse multimodal graph sequential state model (SMGSSM) with multi-constraint model predictive control (MC-MPC). Four neural encoders apply reliability gating to financial and tax data, operational data, cybersecurity disclosures, and Form 10-K semantics. A sparse exposure graph and a gated recurrent unit (GRU) jointly capture interfirm propagation and quarterly temporal dependence, while multitask output heads predict next-quarter resilience and three categories of risk. The control layer allocates liquidity support, supply substitution, cybersecurity hardening, and tax actions subject to budget and adjustment constraints. The key methodological contribution is threefold: separating evidence availability from predictive relevance through reliability gating; jointly modeling interfirm exposure and quarterly memory; and converting predicted cyber-financial-tax states into auditable actions under a shared resource budget. Using a public panel of 24 U.S. manufacturing firms, the calendar-year 2025 holdout root mean squared error (RMSE), after full-sample preprocessing, is 0.0663 ± 0.0089 and increases by 12.1% when graph propagation is removed. Across 120 closed-loop paths, MC-MPC reduces aggregate loss by 34.42% and increases average resilience by 35.90% relative to no control. Compared with reactive rules, it reduces cyber exposure by 41.76%, although production loss and control cost are higher. Within the stated retrospective and simulation-based design, the findings illustrate how cybersecurity can be represented as a controllable component of enterprise resilience and how a shared budget reveals cross-objective crowding-out. They do not establish prospective predictive validity or causal intervention effects.