Traditional, empirical ground‐motion models (GMMs) are developed by prescribing a functional form between predictive parameters and ground‐motion intensity measures. Machine‐learning techniques may serve as a fully data‐driven alternative to widely used regression techniques, as they do not require explicitly defining these relationships. Although, machine‐learning methods offer a nonparametric alternative to regression methods, there are few studies that develop and assess performance of traditional versus machine‐learning GMMs side by side. We compare the performance and behavior of these two approaches: a mixed‐effects maximum‐likelihood (MEML) model and a feed‐forward artificial neural network (ANN). We develop and train both models on the same dataset from southern California. We subsequently test both models on a dataset from the 2019 Ridgecrest sequence, in a new region and on magnitudes outside the range of the training dataset, to examine model portability. Our models estimate horizontal peak ground acceleration, and the input parameters include moment magnitude (⁠M⁠) and hypocentral distance (⁠Rhyp⁠), and some include a site parameter, either VS30 or κ0⁠.We find that, with our small set of input parameters, the ANN generally shows more site‐specific predictions than the MEML model with more variation between sites, and, performs better than their corresponding MEML model, when applied “blind” to our testing dataset (in which the MEML random effects cannot be considered). Although, previous studies have found that κ0 may be a better predictor of site effects than VS30⁠, we found similar performance, suggesting that including a site parameter may be more important than the physical meaning of the parameter. Finally, when applying our models to our Ridgecrest dataset, we find that both methods perform well; however, the MEML models perform better with the new dataset than the ANN models, suggesting that future applications of ANN models may need to consider how to accommodate model portability.

中文翻译：

---

人工神经网络与南加州小震级地震传统地面运动模型的比较

通过规定预测参数和地面运动强度测度之间的函数形式，可以开发出传统的经验地面运动模型（GMM）。机器学习技术可以作为广泛使用的回归技术的完全数据驱动的替代方案，因为它们不需要明确定义这些关系。尽管机器学习方法提供了回归分析方法的非参数替代方法，但是很少有研究可以并行开发和评估传统GMM和机器学习GMM的性能。我们比较了这两种方法的性能和行为：混合效应最大似然（MEML）模型和前馈人工神经网络（ANN）。我们在来自南加州的同一数据集上开发和训练这两个模型。随后，我们在来自 2019 Ridgecrest 序列的数据集、新区域和训练数据集范围之外的震级上测试了这两个模型，以检查模型的可移植性。我们的模型估计水平峰值地面加速度，输入参数包括力矩大小 (⁠M⁠) 和震源距离 (⁠Rhyp⁠)，有些包括站点参数，VS30 或 κ0⁠。我们发现，使用我们的小集合在输入参数中，ANN 通常比 MEML 模型显示更多的站点特定预测，站点之间的变化更大，并且当“盲目”应用于我们的测试数据集时（其中 MEML 随机效应不能考虑）。尽管以前的研究发现κ0可能比VS30⁠更能预测站点效果，但我们发现类似的性能，暗示包含站点参数可能比参数的物理意义更重要。最后，将我们的模型应用于 Ridgecrest 数据集时，我们发现两种方法都表现良好；然而，MEML 模型在新数据集上的表现优于 ANN 模型，这表明 ANN 模型的未来应用可能需要考虑如何适应模型的可移植性。