LinkedIn

LinkedIn

GitHub

GitHub

YouTube

YouTube

Email

YouTube

Resume

YouTube
Data Scientist - Computational Physicist - Researcher

projects.

Long Term Water Pipeline failure prediction using Random Survival Forests

Python

sklearn

xgboost

random forest

ensemble methods

cross validation

survival analysis

Thumbnail
  • Water suppliers incur significant costs for maintaining underground water pipelines. Most of these costs are incurred in performing maintenance and repair of pipes, the failures of which are unpredictable.
  • We set out to suggests a model to be used by three water suppliers in three Australian states: Sydney Water, UnityWater and Western Water, to predict the failure of the pipes.
  • We showed Random Survival Forests to be the best performing paradigm among a host of other random forest methods.
  • Completed in collaboration with Data61 of CSIRO, Sydney Water, UnityWater and Western Water.

Energy demand forecasting using Recurrent Neural Network

Python

pytorch

tensorflow

convolutional filters

RNN

LSTM

neural attention

encoder-decoder

energy demand prediction

energy peak prediction

Thumbnail
  • Electricty demand prediction is a critical problem faced by energy suppliers who need to match the demand with appropriate supply everyday. Daily peak demand is an especially crucial parameter.
  • We performed analyses using various neural network architectures to predict both electricity demand and daily demand peaks.
  • We proposed a recurrent neural network with a novel dual-phase, dual-stage attention mechanism capable of capturing the long and short term effects impacting daily demand peaks.
  • Completed in collaboration with Data61 of CSIRO and Ausgrid.

Rough surface optics prediction fro solar cells using convolutional neural networks

Python

pytorch

data augmentation

convolutional neural networks

densenet

transfer learning

finite element modelling

photovoltaic light management

solar cells

Thumbnail
  • Light management is a key problem in solar cell design. Rough surfaces have shown to aid in trapping sunlight within solar cells, but predicting and understanding this behaviour is difficult due to the complexity of solving the resulting 3D partial differential equations.
  • We proposed to use the power of convolutional deep neural networks to train a model to predict the optical effects at rough surfaces using real atomic force microscopy images of nanoscale surfaces.
  • The densenet neural net was used to perform transfer learning, along with finite element modelling to generate theoretical optical parameters to train the model.
  • Achieved the target error of <1% and currently in the final tuning stages.
  • Complete code will be made available at the conclusion of the project.

Comment classification on the AWS cloud using BERT and word2vec

Python

AWS

data wrangling

machine learning pipeline

BERT

word2vec

Sagemaker

S3

EC2

Thumbnail
  • Text classification is the quintessential natural language processing problem. With the advent of the novel text encoding schemes, text classification has become an easier problem to solve by each passing day.
  • In this project I leveraged the built-in word2vec and BERT encoding schemes as well as simple neural networks to classify product review comments.
  • Machine learning pipelines and other infrastructure available within AWS were used in addition to Sagemaker notebooks.
  • Completed as part of the AWS machine learning specialization offered on Coursera.org.

Efficient solver for the coupled many-emitter cavity quantum model

C++

quantum optics

open quantum systems

coupled quantum systems

OpenMP

sparse matrix multiplication

adaptive Runge-Kutta

Hamiltonian Monte-Carlo

Thumbnail
  • The coupled many-emitter cavity model is a common yet intractable problem in quantum electrodynamics that has received much attention recently due to the interest in plasmonic phenomena.
  • I discovered a novel algebraic symmetry within the quantum equations of motion that allowed me to reduce the computational complexity of the system from quintic to quartic.
  • This allowed for the solution of the system for sizes that have never been possible before.
  • Coded in C++, parallelised using OpenMP and run on supercomputing clusters.

Chromatic analysis of solar cell designs for colored and transparent solar window applications

Python

classical optics

Maxwell's equations'

FDTD

solar windows

nanostructuring

Thumbnail
  • Transparent and colored solar panels have potential for great use as the numbers of high rise buildings with window space grows exponentially all over the world.
  • We perform theoretical calculations on a potential solar cell structure where changing a structural parameter changes the colour seen by people inside and outside a building.
  • This is part of the work done to design solar cells for window applications.
  • Coded in python using the meep FDTD solver to solve Maxwell's equations on a grid.

work.

Research Fellow

Monash University (2020-Present)

Built a full-suite of analytical, computational and machine-learning models to further our understanding of opto-electrical properties of materials in thin-film solar cell technologies.

Teaching Associate

Monash University (2017-2020)

Tutored masters and undergraduate students in Advanced Engineering Data Analysis (ENG5001/6001), Advanced Algorithms and Data Structures (FIT3155), Theory of Computation (FIT2014).

education.

Doctor of Philosophy (Computational Quantum Optics)

Monash University (2016-2020)

Researched theoretical and computational methods for understanding the nanoscale phenomena in plasmonic structures. Computational tools were coded and solved with the aid of supercomputers.

Bachelor of the Science of Engineering (Electronic and Telecommunication Engineering)

University of Moratuwa, Sri Lanka (2010-2015)

Graduated with First class Honours and a GPA of 3.84. Studied Electronic, robotics, algorithms and programming, communication theory, information theory, statistical methods and machine learning.