Is the model capable of forecasting Lorentz system?

[luckyeric320]

I tried the model to performance forecasting on a generated trace of 3-dim Lorentz system but it seemed not work.

[Hy23333]

I tried on flow passing a cylinder mentioned in the paper (data from Steven Brunton), the prediction and error had many 'nan' data, which makes the result very tricky and unreliable. And also, the code did not provide the prediction result in the original space; I mean if you save the prediction data and print the shape, actually the code forget to map the prediction from 64-D into the normal 2-D. I am not sure if this is the reason that the code only show the error curve except the prediction one.

[erichson]

@Hy23333 thanks for reaching out, but your description is a bit vague. The flow past the cylinder is a pretty simple problem and it is relatively easy to setup. Note, that we crop the snapshots a bit, because you have some strange values close to the cylinder which might impact performance if you are not careful about how you normalize the data.

I am not sure what you mean by that the prediction and error has "nan data"? Which code are you referring to? You need to flatten the 2-D (vorticity) snapshots into a vector first. Then you predict the high-dimensional target vectors and then you can reshape the vector into the 2-D snapshot. If you have trouble, you can send me an email and I can share the code with you that I used for this problem.

Best, Ben