deep ensemble: Inconsistent results with a mlp trained with the maximum likelihood principle

i am experiencing an annoying issue with the neural network that i have created and that i use as a core element of a deep ensemble network to learn the uncertainty in a regression task

def NLL(y, distr): 
    return -distr.log_prob(y) 


def normal_sp(params):       
         return tfd.Normal(loc=params[:,0:1], scale=1e-3 + tf.math.softplus(0.005* params[:,1:2]))

def create_model():
 
    inputs = Input(shape=(1,),name="input layer")
    x =  Dense(20, activation="relu",
    kernel_regularizer=regularizers.L2( l2=1e-4),
    bias_regularizer=regularizers.L2(1e-4))(inputs)
  
    x =  Dense(50, activation="relu",
    kernel_regularizer=regularizers.L2( l2=1e-4),
    bias_regularizer=regularizers.L2(1e-4))(x)
  
    x =  Dense(20, activation="relu",
    kernel_regularizer=regularizers.L2( l2=1e-4),
    bias_regularizer=regularizers.L2(1e-4))(x)
  
    params_mc = Dense(2,activation="relu",activity_regularizer=regularizers.L2(1e-5))(x)
    dist_mc = tfp.layers.DistributionLambda(normal_sp, name='normal_sp')( params_mc ) 

    modello = Model(inputs=inputs, outputs=dist_mc)
    return modello 

optimizer = tf.optimizers.Adam(learning_rate=0.0002)

model =  create_model();
model.compile(optimizer=optimizer,loss=NLL,jit_compile=True ) 
model.fit(X, y, epochs=5000, verbose=0, batch_size=batch_size,validation_data=(X_val,y_val) )

the problem is that this network around 30% of the times refuses to learn anything and mess up the entire ensemble. for example this is the plot of the prediction of 5 trained models

the fourth model is stuck around zero and ruins everything

I don't understand what is the problem, it makes no sense.