Failures, successes and challenges of machine-learning-based engineering ground-motion models

The exponential growth of seismological data and machine learning methods offer new perspectives for analysing the factors controlling seismic ground motions and predicting earthquake shaking for earthquake engineering. However, the first models (e.g. Derras et al., 2012) using "simple" neural networks to predict seismic motions did not convince the earthquake engineering community, which continued to use more conventional models. We analyse the weaknesses (from the perspective of engineering seismology) of this first generation of ML-based ground motion models and explain why this first generation did not provide sufficient added value compared to conventional models.  Based on this experience, we propose two evolutions and new methods that have advantages over conventional methods and therefore have greater potential.  A first class of models (e.g. Lilienkamp et al., 2022), based on a U-net neural network, predicts spatial variations in seismic motions (e.g. site effects in three-dimensional basins) by considering seismic motions in map form. A second class of approaches) combines AI methods (conditional generative adversarial networks,  Esfahani et al., 2023) and hybrid databases (observations and simulations selected for their complementarity) to train simulation models capable of generating not only a few parameters (e.g. PGA) describing ground motions, but the full acceleration time histories. We will discuss the potential advantages of this new generation of ML-based methods compared to conventional methods, but also the challenges (and proposed solutions) to best combine simulations and observations, and to calibrate both the best estimate and the variability of future ground motions.