Reservoir Computing for Time Series Prediction Using Spin Torque Oscillators

A proof of concept reservoir computing using a pool of Spin Torque Oscillators (STOs) to predict time series data (Mackey Glass Equation). The STOs are simulated using the Mumax3.

Usage

1. Install Mumax3
2. Clone this repository
3. Generate the .mx3 files using the scripts_generator.py script.
4. Run the simulation (I used a cluster controlled by scripts/reservoir.pbs)

Customize the parameters

Change the number of STOs and the current applied to each STO in the scripts_generator.py file. Change the simulation parameters in the template.txt file.

Post-processing

1. Run main.py to generate the plots and the error metrics.
2. Run plot_final_result.py to generate a plot of the MSE for configurations with different number of STOs.

Results

Predicting the Mackey Glass Equation using 25 STOs. Although the prediction is not perfect, it will be clear when comparing to a naive linear regression model that the reservoir is able to capture the dynamics of the time series data.

Of course it is vital to compare the results without the reservoir computing but just linear regression. Simple ML models usually "predicts" the data by reusing the previous data when dealing with time series data.

The prediction plot might looks good.

But when we zoom in, it is clear that the prediction is just reusing the previous data rendering the model useless.