General info

This work is devoted to numerical study of nanoelectromechanical systems (NEMS) by means of machine learning techniques.

Milestones of the conducted work are listed below.

1. Finite element method (FEM)

In order to obtain data for training machine learning models, FEM simulation results are utilized. In particular, series of NEMS configuration that vary in 8 input parameters (lenght and width of nanobeam, thickness of the 1st and the 2nd layers, temperature, gate voltage, gate distance and mechanical beam pretension) are simulated providing outputs containing 5 crucial parameters (resonant frequency, quality factor, effective mass, thermoelastic losses and mechanical noise spectral density value) of every mode from the 1st to 4th one for given inputs.

For a better control over input parameter distribution, all input parameters were preliminarily generated using code in data/input_parameter_generator/input_generator.ipynb. By means of this code batches of input parameters (each of ~200 elements) were generated and used as input for FEM.

2. Processing of raw dataset

FEM output is processed to a more conventional representation for machine learning tasks by means of code in data/data_formatting.ipynb. Also, anomalies of the raw dataset can be detected and eliminated.

3. Visual inspection of processed dataset

In data/data_review.ipynb input and output parameters distribution in processed dataset are visualized. On rare occasions some anomalies can be detected with the help of these visualizations.

4. Classic machine learning

Models used:

• LinearRegressor (for a baseline)
• RandomForestRegressor
• TabNetRegressor + Optuna
• XGBRegressor + Optuna
• XGBRegressor + custom Cross Validation

Also, custom scaler class was developed.

See details in training/classic/README.md.

5. Neural networks

Models:

• Fully Connected Networks with different architectures
• Loss-wrapped neural network

For details, see training/neural/README.md.