Probabilistic computing utilizing HfO₂-based stochastic ferroelectric tunnel junctions.

Area of Science:

• Materials Science
• Neuroscience
• Computer Science

Background:

• Probabilistic neural networks excel at complex optimization but require specialized hardware.
• Existing solutions for probabilistic bit (p-bit) neurons and artificial synapses often lack energy efficiency and reliability.

Purpose of the Study:

• To develop novel stochastic ferroelectric tunnel junctions (s-FTJs) and reliable ferroelectric tunnel junctions (r-FTJs).
• To implement energy-efficient p-bit neurons and artificial synapses for advanced probabilistic computing.
• To demonstrate the application of these components in solving complex computational problems.

Main Methods:

• Tuning oxygen vacancy concentration in Hf_0.5Zr_0.5O₂ ferroelectric films to create s-FTJs and r-FTJs.
• Utilizing s-FTJs for p-bit neurons capable of tunable probability output and true random number generation.
• Employing r-FTJs for reliable artificial synapses.
• Constructing a hardware prototype of a four-neuron Boltzmann machine for experimental validation.

Main Results:

• Demonstrated s-FTJ-based p-bit neurons with tunable probabilities and low write power (~76 nW).
• Successfully operated the p-bit as a true random number generator at a 0.5 probability.
• The four-neuron Boltzmann machine prototype solved the maximum independent set problem.
• Simulations indicated a 655-neuron Boltzmann machine could accurately predict RNA secondary structure.

Conclusions:

• Ferroelectric tunnel junctions offer a viable pathway to high-performance, low-energy probabilistic computing.
• The developed p-bit neurons and synapses exhibit excellent process compatibility and energy efficiency.
• This technology holds promise for tackling complex optimization and biological sequence analysis tasks.