Total neural network (NN) loss for the solution with λ = 20° and θ_pc = 36° in the closed line (blue) and open line (orange) regions versus training epochs. The differences in epochs arises from the varying number of steps taken by the training algorithm during second-order optimization. Training initially employed a first-order optimizer, which was switched to a second-order optimizer after 10⁴ training epochs. Beyond that, we adjusted the shape of the interim separatrix every 3 × 10⁴ epochs, producing the blue and orange high spikes. The total training loss achieved, between 10⁻⁵ and 10⁻⁶, is considered satisfactory.