Rayleigh wave exploration is a powerful method for estimating near-surface shear-wave (S-wave) velocities, providing valuable insights into the stiffness properties of subsurface materials inside the Earth. The dispersion curve inversion of Rayleigh wave corresponds to the optimization process of searching for the optimal solutions of earth model parameters based on the measured dispersion curves. At present, diversified inversion algorithms have been introduced into the process of Rayleigh wave inversion. However, limited studies have been conducted to uncover the variations in inversion performance among commonly used inversion algorithms. To obtain a comprehensive understanding of the optimization performance of these inversion algorithms, we systematically investigate and quantitatively assess the inversion performance of two bionic algorithms, two probabilistic algorithms, a gradient-based algorithm, and two neural network algorithms. The evaluation indices include the computational cost, accuracy, stability, generalization ability, noise effects, and field data processing capability. It is found that the Bound-constrained limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS-B) algorithm and the broad learning (BL) network have the lowest computational cost among candidate algorithms. Furthermore, the transitional Markov Chain Monte Carlo algorithm, deep learning (DL) network, and BL network outperform the other four algorithms regarding accuracy, stability, resistance to noise effects, and capability to process field data. The DL and BL networks demonstrate the highest level of generalization compared to the other algorithms. The comparison results reveal the variations in candidate algorithms for the inversion task, causing a clear understanding of the inversion performance of candidate algorithms. This study can promote the S-wave velocity estimation by Rayleigh wave inversion.

瑞利波探测是一种估算近地表剪切波（S 波）速度的强大方法，可为地球内部地下材料的刚度特性提供有价值的见解。瑞利波频散曲线反演相当于根据实测频散曲线寻找地球模型参数最优解的优化过程。目前瑞利波反演过程中已经引入了多种反演算法。然而，为揭示常用反演算法之间反演性能的差异而进行的研究非常有限。为了全面了解这些反演算法的优化性能，我们系统地研究并定量评估了两种仿生算法、两种概率算法、一种基于梯度的算法和两种神经网络算法的反演性能。评价指标包括计算成本、准确性、稳定性、泛化能力、噪声影响和现场数据处理能力。研究发现，有界约束的有限内存 Broyden-Fletcher-Goldfarb-Shanno (L-BFGS-B) 算法和广泛学习 (BL) 网络在候选算法中具有最低的计算成本。此外，过渡马尔可夫链蒙特卡罗算法、深度学习（DL）网络和BL网络在准确性、稳定性、抗噪声影响和处理现场数据的能力方面优于其他四种算法。与其他算法相比，DL 和 BL 网络展示了最高水平的泛化能力。 比较结果揭示了反演任务的候选算法的变化，使人们能够清楚地了解候选算法的反演性能。该研究可促进瑞利波反演横波速度估算。