Simulink's Stateful Predict block produces results that diverge from MATLAB's predict function when using the same trained network

Hello,

I trained an LSTM network to predict the contact force in a pneumatic polishing process. In MATLAB, the network's predict output using 'predict' function matches the measured force well, but when I deploy exactly the same network inside a Simulink 'Stateful Predict' block the predicted signal drifts and exhibits completly different result, as shown in the attached figure 1.

I attempted resolving this issue by adjusting Stateful Predict block settings and setting Simulink step time to match that used for neural-network training. However, despite these attempts, the results were always the same as in figure 1.

Questions

1. What checklist or recommended workflow ensures identical results between MATLAB predict and Simulink Stateful Predict for the same network and inputs?

2. Is this discrepancy a known bug or something slated for a future patch?

Figure 1.

(Left) Results using Matlab predict function // (Right) Results using Simulink Stateful Predict block

Figure 2.

Simulink configuration state

[Pratyush Swain]

Hi, can you share your model & script for reproducing the issue or atleast trimmed down version of it which still shows the discrepancy.

Thanks

[Bahadir]

did you find any solution, ı have same problem.