FAQs

FAQs:

1. Why the sequence of reactions learned can be random? Does it matter?

   The learning of CRNN involves stochastic gradient descent optimization. Therefore, the learned sequence of reactions could be random and is different from the one presented in the paper. However, the sequence of reactions doesn't matter for modeling and prediction.

2. How many experiments are required for the learning of CRNN?

   This is a very heuristic question. It is worthy to note that we shall analyze the training of CRNN from both the perspective of fitness and parameter inference. By default, one will pay attention to the fitness by inspecting the loss function. However, the goal of learning CRNN is not only to fit the available data but also to narrow down the model parameter uncertainties. Besides checking the convergence of loss functions, one should also check the convergence of model parameters. There are many potential ways to do it, such as bootstrapping, bayesian inference, bayesian neural networks.

3. What ODE solver I should choose?

   For non/mild-stiff systems, the default choice would be Tsit5(). For a strong stiff system, try AutoTsit5(Rosenbrock23()), or TRBDF2() for larger system.

4. Can I apply CRNN to very-stiff systems?

   Currently, our experience is that the standard approach for learning neural-ode for stiff systems is very challenging. It applies to either standard neural networks or CRNN. We are actively working on learning CRNN for very stiff systems, and there are some successful demonstrations in some systems. For example, one of the classical strong stiff systems, Robertson's problem. You can find the preliminary results on Robertson's problem in this repo.