Design and Applications of Bio-Molecular Memelements in Neuromorphic Computing
Restricted (Penn State Only)
- Author:
- Mohamed, Ahmed
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- January 25, 2025
- Committee Members:
- Robert Kunz, Professor in Charge/Director of Graduate Studies
Herschel Pangborn, Major Field Member
Joseph Najem, Chair & Dissertation Advisor
Jean-Michel Mongeau, Major Field Member
Abhronil Sengupta, Outside Unit & Field Member - Keywords:
- Memristor
Memcapacitor
Nonlinear Dynamics
Reservoir Computing
Artificial Intelligence
Neuromorphic Computing
Short-term Memory
Machine Learning
Neuristor
Mem-Impedence - Abstract:
- The increasing energy demands of artificial intelligence (AI) systems, particularly in energy- and space-constrained applications like implantable medical devices and remote sensors, highlight the need for more energy-efficient computing paradigms. Drawing inspiration from the human brain's unparalleled energy efficiency, this dissertation explores using soft biomolecular components, i.e., synthetic lipids and peptides, to develop neuromorphic computing systems. We focus on biomembrane-based memcapacitors and ion-channel-based memristors for physical reservoir computing (PRC) applications, mimicking synaptic functionalities. In addition to synaptic emulation that mimics learning and memory, we developed a model for the neural component—a Hodgkin-Huxley-inspired biomolecular neuristor guided by Chua’s Local Activity Principle—completing the framework for generating and propagating action-potential-like signals. In the first chapter, we include the motivation behind this work and impart the reader with the essential background information to understand and appreciate its significance. In the second chapter, we introduce a homogeneous PRC using biomembrane-based memcapacitors for complex tasks with ultra-low energy consumption. With this architecture, we demonstrated marked prediction for synchronous and asynchronous classification and time series prediction tasks. In the third chapter, we leverage the intrinsic voltage offsets from distinct memcapacitor compositions to design a heterogeneous reservoir that enhances high-dimensional mapping via paired-pulse facilitation and depression. This approach reduces pre-processing overhead and outperforms traditional input-encoding approaches. In the fourth chapter, to further increase computational density, we develop a mem-hybrid reservoir combining memcapacitors and memristors. This dual-function device provides distinct memory and dynamical properties, allowing for task-specific optimization and, more importantly, addressing the memory-nonlinearity trade-off in PRC. In the fifth chapter, inspired by neuronal action potentials, we introduce an ion-channel-based neuristor model by integrating two ion-channel-based memristors in anti-parallel, mimicking sodium and potassium channel dynamics. This neuristor model closely emulates neuronal firing generation and propagation in biological axons and bridges the gap between synaptic and neuronal processes. The outcomes of this research advance energy-efficient, adaptable neuromorphic systems for environments where traditional electronic substrates are unsuitable and offer novel insights on nonlinear electronics design for future AI implementations.