My take on this is that noise currents from the VFD's are producing noise voltages on the panel buses for each of the three phases. As a result the noise voltage will be proportional to the impedance presented at each bus. And this impedance is likely be higher at the frequencies significantly above 60 Hz where VFD generated noise is present.
However, there will likely be relatively little noise on the neutral bar because it is not connected to the VFD's. And so a noisy line voltage and a quiet neutral will be fed to the GFCI receptacles. Hence the line and neutral pair of conductors will have both a differential component and a common mode component of noise, such that these noise components effectively add on the line conductor and subtract out on the neutral. The bottom line of this is that there will be a common mode noise component that is fed to to GFCI receptacles.
The question is whether a GFCI receptacle will be false triggered by this noise voltage without a load being present, or if some load impedance is needed to produce a noise current that will be sensed by the internal toroid and cause it to trigger. This is why I asked above whether there were loads on the GFCI receptacles.
Hence it might be useful to see whether the GFCI receptacle will trip without any load. If it doesn't trip when there is no load, then a 2-wire load like a heater could be plugged in which would present a very high impedance to a common-mode noise voltage present on the 2 wires. If it still doesn't trip then that would indicate that a common-mode impedance to the EGC is needed (for example, from a capacitor) in order for a common-mode noise current to flow and then trip the GFCI.
As far as mitigating measures, ferrite cores could be put around the L-N pairs to see if they might reduce common-mode noise sufficently to prevent the tripping.
