Project Supervised by

• Prof. (Dr.) Tarun Kanti Bandopadhyay
• Dr. Ryan Gosselin

Problem Statement

• Finding Reynolds Number Equivalent for Mixing Powders in a feedframe
• Learning DEM Particle Behaviour from data
• Optimising CFD data using ANN

Context

Feedframe

• Discrete Element Modelling of Feedframe
• Calculating Danckwert's Segregation Intensity
• Plotting DI w.r.t time and with other factors to find a pattern common to all

DEM Particle Behaviour from data

• With enough training sets, can a deep learning algorithm essentially recreate an entire simulation without appreciable loss of accuracy and precision ?

Aims and objectives

FeedFrame

• Plotting SI vs Time Vs Factor

Factors

• Angular Velocity
• Radius Ratio
• Density Ratio
• Inclination Angle
• Geometric Manipulation
• Volume Percentage of Particles
• Sinusodial rotation (Frequency)

Plot

Interpolating the results to generate a 3D surface plots using appropriate methods

DEM Learning

• Creating a Deep Network that can predict the time series of a DEM Simulation
• The whole simulation of DEM can be concatenated into a matrix which will have the following format
  • P x f x T

Here P is the number of particles, f is the number of features ( Positions, Velocity and Force) and T is the timesteps.

Time Series prediction using Neural Network

Neural Network with Memory

Tasks

Progress on project so far

Feedframe

• [75%] Building a computational pipeline to calculate the simulation
  • [X] Using LIGGGHTS as a shared libary
  • [X] Creating input structure
  • [ ] Taking snapshots of the simulation
  • [X] PyDoIt Make tool to automate tasks

• [40%] Use of Design of Experiment for Factorial Design
  • [X] Choice of Factors
  • [X] Selection of ranges
  • [ ] Specific Levels at which runs will be made
  • [ ] Selection of Response Variables
  • [ ] Choice of Experimental Design
• [ ] Performing the experiment
• [ ] Data Analysis
• [ ] Conclusions

DEM Learning

• [50%] Literature Review
  • [X] Learning about RNNs and Restricted Boltzman Machines
  • [ ] Learning more about Deep Learning and hyperparameter optimization
• [ ] Training an RNN on a smaller dataset
• [ ] If the training seems successful, then use of cluster to train the feedframe model

Problems

• Require more computational power approx 128 GB RAM to train RNNs which take a lot of time ( 3 days for 200K datasets ) to train
• Can rent GPUs for very cheap per hour price

Conclusions

• Deep Learning is a vast field and which is changing/improving at a tremendous rate( which is a good thing for mankind)
• Using Design of Experiments will provide a good way to detect any sort of phase change in the mixing

Thank you