Create "PredictorFunction" with NeuralNetwork

Question

I tried to construct a predictor for a continuous variable using neural networks, something like

f(a,b,c,d,e) = x,

trying to predict x for a given set of a,b,c,d, and e.

The builtin Predict[] already did a great job, running

f = Predict[trainingData, Method->"NeuralNetwork"].

Because in the end this is just for prototyping, and I want to rebuilt the network "manually", I extracted the layer information, and it looks like this:

Afterwards I created a NetChain to reproduce the results from the PredictorFunction

net = NetChain[

   Flatten[
    {LinearLayer[50],
     
     Table[{
       LinearLayer[50],
       ElementwiseLayer["SELU"],
       DropoutLayer[0.5]
       }, {i, 
       10}],
     LinearLayer[50], LinearLayer[1]}]
   
   , "Input" -> 5
   ];

The problem arises here: After training, my self-made network performs significantly worse (always predicting a value very close to the mean of the training data set), so I guess the training process is significantly different.

result = NetTrain[ net, trainingData, All, MaxTrainingRounds -> 20]
f = result["TrainedNetwork"]

Is there a significant difference in the training process, if NetTrain is used rather than Predict[], or is something else causing the bad performance of my network?

Thanks in advance