NEST stalling for specific sets of parameters of a NEST ML model

[DrTaDa]

Hello, When running the NEST ML version of the aeif_cond_exp neuron model, NEST "stalls" during the nest.Simulate(100.) command for specific sets of parameters. This does not happen when using the same parameters in the NEST aeif_cond_exp neuron model.

The issue can be reproduced with the following code and the attached NEST ML model (aeif_cond_exp_test_neuron.nestml.zip)

import nest

nest.Install("nestmlmodule")

params = {
        'E_L': -80,
        'V_reset': -60.0,
        'refr_T': 2.0,
        'g_L': 4,
        'C_m': 32,
        'E_exc': 0.0,
        'E_inh': -80.0,
        'tau_syn_exc': 1.5,
        'tau_syn_inh': 4.2,
        'a' : -0.8,
        'b' : 80,
        'Delta_T' : 0.8,
        'tau_w' : 144.0,
        'V_th' : -57.0                          
    }

neuron = nest.Create("aeif_cond_exp_test_neuron", 10)

for param in params:
    neuron.set({param: params[param]})

neuron.set({'v': -54})

nest.Simulate(100.)

Could you let me know if there is a way to debug NEST ML models ? Or If there is something obvious I am missing ?

nest-simulator==3.5 nestml== 7.0.2.post0.dev0 (also tried with 8.0.0rc1.post0.dev0) cmake version 3.16.3

[clinssen]

Hi, thanks for writing in! I am not an expert on the model, but is it okay for the parameter $a$ to take on negative values? In case this makes the system mathematically unstable, it might cause issues with the numerical integrator.

From Izhikevich (2003):

The parameter a describes the time scale of the recovery variable u. Smaller values result in slower recovery. A typical value is a = 0.02

[DrTaDa]

Hello @clinssen, thank you for your answer. Sorry, I left our custom parameter set. It does not work either with the value of 0.02.

I was more curious about why it works when using the NEST model but not the NEST ML model. The equations for both being identical. Is there a way to debug NEST ML models ? Here the model does not even pass the first time step.

[clinssen]

Thank you for the update! I believe this is due to NEST having more stringent solver tolerances than NESTML by default. I have updated the defaults in #1086. Could I ask you to try the code from that PR to see if it solves your issue? Much obliged!

For more debug output, you could uncomment the #define DEBUG flag in the C++ jinja2 template files, and you can also add your own print() statements directly in the NESTML files (or directly in the C++ code using std::cout).

[DrTaDa]

It works ! Thank you for your help, you are a lifesaver.

[clinssen]

Super cool, no prob! Please let us know if you run into any further issues.