Dynamic Non-equilibrium Supervisor for Real-Time Weight Updates and Non-linear Memristor Homeostasis

[Zhenis-ref]

Hello MemTorch team and fellow researchers,

I have read your publication in Neurocomputing with great interest. The MemTorch framework's capability to co-simulate device non-idealities and peripheral circuits is an outstanding contribution to the neuromorphic computing community.
In most current Memristive Deep Learning Systems (MDLS), stochasticity, ionic drift, and device aging are treated as destructive "non-idealities" that must be suppressed or heavily calibrated via complex digital control overlays.
I would like to share an alternative perspective that we are currently developing within the framework of non-equilibrium dynamic systems (the ΔN–ΔD theoretical model). We investigated a fundamental question: instead of suppressing these inherent physical non-idealities, can we utilize them as an endogenous driver for self-organization and structural homeostasis?
Rather than trying to control the weight of each individual memristor sequentially, our approach introduces a macro-supervisor. It treats the structural tension of the matrix (ΔD) as an emergent, measurable property and responds with a unifying external balancing force field (U) while dynamically adjusting the system's viscosity (α).
Through numerical experiments (implemented in our open-source framework, DN-Memristor-Sim v0.6), we demonstrated that under critical external stress, the analog substrate of the crossbar undergoes a non-linear phase maneuver. It transitions into a stable sparse switching regime, automatically preserving structural parameters without relying on conventional gradient descent or high-overhead digital tuning. Conversely, standard memristive substrates under identical stress scenarios rapidly degrade into uncontrolled switching avalanches or saturated conduction paths.
We have published our reproducible simulation platform on GitHub, fully frozen and verified via Zenodo:
👉 Zhenis-ref/DN-Memristor-Sim (DOI: 10.5281/zenodo.21206728)
I would be highly interested in opening an academic discussion on the following points:

1. Do you see potential in integrating macro-dynamic supervisors on top of MemTorch crossbars to protect DNN weights from degradation under dynamic environments?
2. Have you encountered any phenomena of spontaneous emergent stabilization in your simulations when the physics of the crossbar is allowed to evolve without rigid digital constraints?
   Looking forward to your valuable insights!

Best regards,
Zhengis Tileubay
Independent Researcher in Systems Theory and AI
Astana, Kazakhstan