Existing Intelligent Reflectible Surface (IRS)-assisted communication systems in which only angular and delay parameters are taken into account for channel modeling array response cause significant modeling accuracy impacts in systems with wideband large-array IRS. In this regard, the idea of beam squint effect in Multiple-Input Multiple-Output (MIMO) systems is borrowed, and its evolution is applied to IRS-assisted communication systems to obtain a channel modeling formulation applicable to wideband large array IRS systems. In addition, the current channel estimation schemes in IRS-assisted communication systems suffer from the problem of a single incomplete channel matrix for estimation, and this paper constructs an efficient channel estimation protocol to improve the number of channels for a single channel estimation by the one-to-one design of Orthogonal Frequency Division Multiplexing (OFDM) signals and IRS phase-shift modes for subsequent analysis and use. However, realistic signal reception suffers from residual Carrier Frequency Offset (CFO), which have a significant impact on channel estimation, for which the paper improves the estimation accuracy by introducing a Deep Learning (DL) algorithm. The simulation results show that a channel modeling approach for large array IRS wideband systems that takes into account the beam squint effect is necessary; the DL algorithm can effectively deal with the algorithmic performance degradation caused by CFO as compared to conventional algorithms.