Fourier Neural Operators for Accelerating Earthquake Dynamic Rupture Simulations

ABSTRACT Dynamic rupture modeling plays a crucial role in unraveling earthquake source processes. However, the multiscale nature of rupture propagation pose significant challenges, and classical numerical methods remain computationally expensive. To overcome this hurdle, we present a methodology that is both computationally efficient and quantitatively accurate. Specifically, we introduce a surrogate model, in the form of a Fourier Neural Operator, for emulating the nonlinear equations governing dynamic rupture propagation on frictional interfaces. This surrogate is trained on synthetic data generated by multiple physics-based dynamic rupture simulations and is then applied to unseen problems. The proposed methodology retains the accuracy of traditional multiscale methods at a significantly reduced computational cost, achieving a speedup of up to 400,000 compared to the state-of-the—art conventional methods. We evaluate this approach using various examples and demonstrate its efficacy in capturing the spacetime evolution of fault slip rates for a wide range of stress conditions. This development advances the state of the art of computational earthquake dynamics and opens new opportunities for accelerating physics-based rupture forecasts.摘要 动态破裂模拟是解析地震震源过程的关键手段。然而，破裂传播的多尺度特性带来了显著研究挑战，且经典数值方法的计算成本始终居高不下。为突破这一技术瓶颈，本文提出了一种兼具计算高效性与定量准确性的研究方法。具体而言，我们构建了以傅里叶神经算子（Fourier Neural Operator）为架构的替代模型，用于模拟控制摩擦界面动态破裂传播的非线性控制方程组。该替代模型基于多组基于物理机理的动态破裂模拟生成的合成数据集进行训练，随后可直接应用于未见的测试问题。所提方法在保留传统多尺度方法精度的前提下，大幅降低了计算开销，与当前前沿的传统数值方法相比，最高可实现40万倍的加速比。我们通过多类典型算例对该方法进行了验证评估，证明其能够在广泛的应力条件下，准确捕捉断层滑动速率的时空演化特征。这一研究进展推动了计算地震动力学领域的技术前沿，并为加速基于物理机理的破裂预测开辟了全新的机遇。