Reduced-order modelling of Cascadia’s slow slip cycles

Slow-slip events (SSEs) modulate the earthquake cycle in subduction zones, yet understanding their physics remains challenging due to sparse observations and high computational cost of physics-based simulations. We present a scientific machine-learning approach using a data-driven reduced-order modeling (ROM) framework to efficiently simulate the SSE cycle governed by rate-and-state friction in a Cascadia-like 2D subduction setting. Our approach projects fault slip, slip-rate, and state variable trajectories onto a spline-based latent space, which is subsequently emulated using proper-orthogonal decomposition and radial-basis-function interpolation. Achieving a speedup of ~360,000 compared to volumetric simulations, the ROMs enable comprehensive parameter exploration and Bayesian Markov chain Monte Carlo (MCMC) inversion. Our analysis reveals complex, non-linear dependencies of SSE characteristics on the width and magnitude of the deep, low-effective-normal-stress region. Our MCMC inversion constrained by Northern Cascadia SSEs observations indicates near-lithostatic pore fluid pressure (99.6±0.17% lithostatic) and positions the upper frictional transition zone at 30.4 ± 2.8 km depth, consistent with geophysical observations. The inversion resolves the deep SSE-portion of the slab spanning 45±16 km with low effective normal stress of 3.8±1.4 MPa. This framework provides a new tool for advancing the physics-based understanding of SSEs and subduction zone faulting mechanics. By systematically linking megathrust properties such as fluid pressure and fault strength to rate-and-state friction governed slow slip cycle characteristics, such as recurrence interval, our approach helps to constrain the first and second-order physics-based controls and the uncertainties of how plate boundaries slip.