sciml_train was removed a few years ago when Optimization.jl was created as a complete optimization package. In 2020 sciml_train was created in DiffEqFlux as a SciML-focused training system that mixed BFGS methods with neural network methods (ADAM), and by 2021 that system became a full-fledged optimization library Optimization.jl. So sciml_train was deprecated and the deprecation message I think was removed in late 2023.
I recommend looking at the tutorial that I linked in this section:
which shows how to use the Optimization.jl package for doing the optimization process. You define an OpimizationProblem with the loss function (uses the same definition as you have already), and then call solve on the optimization problem using the same algorithms as you had before, so the update should be rather smooth.