Chaotic systems exhibit complex spatiotemporal dynamics that are notoriously difficult to characterize, limiting their applications in fields such as secure communications. To address this challenge, we propose the Chaotic SpatioTemporality-Informed Neural Network (CSTI-NN), a physics-conscious artificial intelligence (AI) framework designed to directly capture the intrinsic spatiotemporality of high-dimensional chaos. By deconstructing and deciphering the Ikeda delay dynamics, our approach successfully reveals previously hidden chaotic invariants that govern correlation scaling. The inherent stability of these invariants against noise directly endows the framework with its enhanced noise-resistant capabilities. In the experiment of the optical chaotic communication, this method achieves over 90% accuracy in reconstructing chaos from noisy signals, surpassing the performance of conventional methods by ∼ 3 dB in signal-to-noise ratio (SNR) tolerance. This endogenous spatiotemporality enables high-quality synchronization, leading to a 10-fold reduction in the bit error rate (BER) at a rate of 5 Gbaud. This work establishes a new paradigm in which AI can not only predict complex behaviors but also fundamentally uncover the underlying physics of chaos, opening new avenues for the development of noise-immune information systems