In the past before I use machine learning, I would create several calculated columns out of the tabular inputs. For example I will calculate the squares and reciprocals, and the multiplication cross different columns. Do I need to do the same before I call the fit function? Or the neurons should be able to catch the possible interactions and the nonlinear relationship among the inputs?
Honestly, here’s a great chance to do some experimentation and see if feature engineering is needed. I’d run the model without any engineering to start, see how your metrics look, then try it again with those calculated columns, cross columns, etc, and see if it helps.