What I did

• Monday: NN data-loading
• Tuesday: NN pre-processing
• Wednesday: NN pause. I need to test that my gaussian averaging method is valid enough to be used for training data
• Thursday: Generated real disordered particles and ran simulations
• Friday: Made particle averaged xanes of real disordered structure and compared it to gaussian averaged method. Still thinking about the results. I also create a github organization for my Brookhaven National Lab group and applied to two jobs today.
• Saturday and Sunday: Applied to jobs and did some SQL practice

What I learned

• Monday: I spent awhile thinking about data-loaders, but my dataset is all text. So there’s no real reason I shouldn’t just load everything into memory at once.
• Tuesday: In tensorflow, you can add a first layer of your NN that will cale0transform the input data in the same way each time. Solves data-leakage problems. It’s a cool solution
• Wednesday: I Frankensteined some old code I wrote to shift particles in random directions via a distribution with the code I recently wrote to generate many structures. The way it works is incredibly convoluted, but at the same time, quite elegant.

def gen_delta_rho_shift(df, shift_sigma):
  df_temp = df.copy()
  df_temp['shift_val'] = np.random.normal(loc=0, scale=shift_sigma, size=df_temp.shape[0])
  df_temp['theta'] = 6.28*np.random.random_sample(df_temp.shape[0])
  df_temp['phi'] = 6.28*np.random.random_sample(df_temp.shape[0])
  # shift_val=radius of sphere project new point onto = distance of new disordered atom from original location
  df_temp['x'] += round(df_temp.shift_val*np.sin(df_temp.phi)*np.cos(df_temp.theta), 5)
  df_temp['y'] += round(df_temp.shift_val*np.sin(df_temp.phi)*np.sin(df_temp.theta), 5)
  df_temp['z'] += round(df_temp.shift_val*np.cos(df_temp.phi), 5)
  # turn to numpy array
  x1 = df_temp.loc[:, 'x'].values
  y1 = df_temp.loc[:, 'y'].values
  z1 = df_temp.loc[:, 'z'].values
  return x1, y1, z1

I’ve been thinking about this code, and I think there’s a relationship between the middle section of that code and the dot product with the unit vector. I could probably greatly improve the code’s efficiency if I figure this out.

• Thursday: Simulations complete. Ran them on the BNL cluster to save time.
• Friday: Learned about a software dev job and talked about what “agile” and “scrum” environments are actually like. Unclear if gaussian averaging method will actually work.

What I will do next

• Create new realistic disordered structures and compare to the gauss averaged structures