An interpretable error indicator for neural-network surrogate models based on encoder–regression decomposition