Neural Compact Modeling of RRAM Conductance for LTP and LTD
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing equation-based compact models for resistive random access memory (RRAM) struggle to accurately predict long-term potentiation (LTP) and long-term depression (LTD) characteristics due to the variety of materials and complex conduction mechanisms, limiting their effectiveness in neuromorphic systems.
Innovation Solution
A data-driven neural network-based compact modeling method using a gated recurrent unit (GRU) cell and multilayer perceptron (MLP) to predict conductance changes in RRAM, incorporating a fully connected layer for instantaneous state changes and training with irregularly sampled input pulse voltages.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If equation-based compact models are used for RRAM, then the model structure is simple and easy to implement, but the prediction accuracy for LTP and LTD characteristics deteriorates due to material variety and complex conduction mechanisms
Solution Approach 1:
The patent replaces traditional equation-based mathematical models with a neural network-based data-driven model. The neural network learns complex non-linear relationships from experimental data, substituting the need for explicit mathematical equations that struggle to capture the complexity of RRAM conduction mechanisms and material variations.
Solution Approach 2:
The patent transforms the modeling approach from fixed equation parameters to adaptive neural network parameters that are learned from data. The model inputs include voltage, time, and previous conductance values, while the neural network dynamically adjusts its internal parameters (weights and biases) to accurately predict LTP and LTD characteristics across different materials and conditions.
2Adaptability or versatility
If traditional neural network models are used, then data-driven flexibility is improved, but the ability to accurately predict both LTP and LTD characteristics simultaneously deteriorates
Solution Approach 1:
The patent implements a dynamic neural network model that processes sequential data with time-dependent behavior. The model uses recurrent connections to capture the temporal evolution of conductance changes, allowing it to dynamically adapt to different programming conditions and accurately predict both LTP (long-term potentiation) and LTD (long-term depression) characteristics that occur at different time scales.
Solution Approach 2:
The patent extends the neural network model to handle multi-dimensional input space including voltage magnitude, pulse width, time intervals, and previous conductance states. By incorporating multiple input dimensions and using a multi-layer architecture, the model can simultaneously capture the complex interactions governing both LTP and LTD phenomena.
3Device complexity
If equation-based models are used, then computational simplicity is maintained, but the ability to handle non-uniform time steps and continuous dynamics deteriorates
Solution Approach 1:
The patent replaces traditional time-step-based numerical methods with a neural network that naturally handles continuous time inputs. The model accepts non-uniformly sampled time data and learns the continuous temporal dynamics directly from the data, eliminating the need for fixed time step discretization and associated computational complexity.
Data Source
AI summary
Disclosed is a compact modeling method for a memory using a neural network performed by a processor. The compact modeling method for a memory using a neural network includes updating a hidden state at time tn+1 by applying a voltage of the memory, a conductance of the memory, and the hidden state approximated at time tn+1 to a gated recurrent unit (GRU) cell, and predicting a conductance of the memory at time tn+1 by applying the voltage, the conductance, and the updated hidden state to a multilayer perceptron (MLP).