Financial market simulation (FMS) serves as a promising tool for understanding market anomalies and underlying trading behaviors. To ensure high-fidelity simulations, it is crucial to calibrate the FMS model to generate data closely resembling the observed market data. Previous efforts primarily focused on calibrating merely the mid-price data, leading to essential information loss of the market activities and thus biasing the calibrated model. The limit order book data (LOB) is the fundamental data that fully captures the market microstructure and is adopted by exchanges worldwide. However, LOB is not applicable to existing calibration objective functions due to its unique properties, including high dimensionality, differences in feature scales, and cross-feature constraints. This article proposes an efficient Transformer-based autoencoder to explicitly learn the vectorized representations of LOB, addressing its unique challenges. The resulted latent vector, which captures the major information of LOB, can then be applied for calibration. Extensive experiments show that the learned latent representation preserves not only the nonlinear auto-correlation in the temporal axis but also the precedence between successive price levels of LOB. Besides, we verify that the representation learning aligns with the downstream calibration tasks, making the first advancement in calibrating FMS on LOB.