Stability criteria for seismic data interpolation artificial neural networks

The widespread applications of artificial neural networks in computational science have drawn deep concerns to their vulnerability to small perturbations. Artificial neural networks are also widely applied to solving various geophysical problems, which should also suffer from the stability issue. We propose two criteria to assess the stability of neural networks that are applied in seismic data interpolation problems: tiny worst-case perturbations and small structural recovery. Quantitative analysis of three artificial neural networks show that instability is commonly observed in all networks and our proposed metrics are reliable to quantify and compare different networks' stability.