Time delay neural networks reveal pressure-independent fault rupture processes in laboratory acoustic emission

Fault nucleation and growth are central to earthquake hazard. Here we analyse fault development in Alzo granite under triaxial confining pressures of 5–40 MPa using a time-delay neural network applied to multi-parameter acoustic-emission data. The model integrates waveform attributes (peak delay, scattering) with occurrence metrics (event rates, Gutenberg–Richter b-value, spatial fractal dimension) to track the transition from distributed microcracking to localised faulting. Genetic algorithms optimise the network, which dynamically weights parameters to characterise fault growth. We find three phases consistent across pressures: microcrack nucleation marked by scattering changes; fault growth captured by evolving spatial and magnitude distributions; and coalescence with rapid peak delay increases and b-value change. The model predicts stress-drop timing and size across pressures and failure mechanisms, from axial splitting to shear localisation, linking waveform features to physically interpretable phases of deformation. Three distinct phases of fault evolution are identified in acoustic emission data from triaxial deformation experiments of Alzo granite evaluated using a time delay neural network optimised via genetic algorithms.