Electrochemical energy storage is becoming one of the most promising solution for renewable energy integration. Liquid metal battery is a prospective battery chemistry for stationary energy storage due to its low cost and long lifespan. However, the flat voltage platform and low working voltage easily introduce relative errors, resulting in challenges in battery state of charge (SOC) estimation. Meanwhile, practical applications of liquid metal batteries require efficient SOC estimation algorithms for massively parallel computing. Thus, in this paper, an improved sliding mode observer (ISMO) is proposed for liquid metal battery SOC estimation to meet the challenges. Firstly, based on a combined equivalent circuit model, the forgetting factor recursive least square algorithm is utilized to identify model parameters in the whole working range. Secondly, a direct differentiation method is put forward to deal with the linearization between the open circuit voltage and the SOC. Finally, a novel adaptive law is proposed to accelerate the convergence, restrict the probable large chattering and improve the estimation accuracy of the algorithm. Compared to the conventional model-based methods, the proposed ISMO exhibits faster convergence, higher accuracy, stronger robustness and lower computational cost in simulations, which indicates an industrialization prospect.