Ira Shokar

PhD Student in Mathematics | University of Cambridge

About

I am a PhD student at the Department of Applied Mathematics and Theoretical Physics (DAMTP) at the University of Cambridge as part of the UKRI Centre for Doctoral Training in the Application of Artificial Intelligence for Environmental Risk and a resident member of Pembroke College.

I am supervised by Professors Peter Haynes FRS and Rich Kerswell FRS as part of the Atmosphere-Ocean Dynamics research group.

My research interests lie in:

  • Probabilistic deep learning for emulation and forecasting
  • Atmospheric and oceanic dynamics
  • Nonlinear, chaotic and stochastic dynamical systems


My current research is focused on utilising deep learning to develop emulators that speed up simulations in fluid mechanics, specifically in cases where a system is under-resolved, requiring probabilistic models to represent uncertainties and produce ensembles. I am interested in deep learning models that can generalize across different parameter regimes while also ensuring that they faithfully replicate the physical properties of the target system.

Research

Please see my Google Scholar profile for an up-to-date list of publications.

Academic Research

Modeling 3D Chaotic Dynamics with Probabilistic Deep Learning - Research Article (Ongoing Research)

2024

  • We capture chaotic dynamics in an under-resolved multi-layer fluid flow, using a spatial autoencoder and stochastic temporal attention to evolve the latent dynamics autoregressively, extending the Stochastic Latent Transformer to 2D and 3D inputs.
  • We examine the statistical equivalence between long-term emulations and numerical integrations and the model’s efficacy in capturing spontaneous transition events exhibited by the system.
  • Supervisors- Professors Peter Haynes and Rich Kerswell.
  • Python: PyTorch; Julia ; LaTeX - Repository

Generalising across regimes to assess extreme events - openreview.net/forum?id=7a5gUX4e5q Research Article (ICLR 2024 Workshop on AI4DifferentialEquations In Science)

2024

  • This study employs the Stochastic Transformer pre-trained on a singular parameter regime, utilising fine-tuning on a significantly smaller dataset of a stochastically forced nonlinear partial differential equation (SPDE).
  • Diverse phenomena emerge in different regimes, yet the Stochastic Transformer adeptly captures the varied dynamics, even when interpolating to unseen regimes.
  • The model is able to successfully capture the spontaneous transition events exhibited by the system across the parameter range.
  • Supervisors- Professors Peter Haynes and Rich Kerswell.
  • Python: PyTorch; Julia ; LaTeX - Repository
  • Stochastic Latent Transformer Generalisation

    Latitude-time plots of stochastic beta-plane turbulence showcasing the ability of the Stochastic Latent Transformer to generalise across various regimes of a key parameter β, values of which were unseen during training. The Stochastic Latent Transformer yeilds results five orders of magnitude faster than conventional numerical integration methods.

Stochastic Latent Transformer - doi:10.48550/arXiv.2310.16741 Research Article (Under review).

2023

  • A probabilistic machine learning (ML) approach for modelling the time evolution of stochastically driven systems.
  • Applying this approach to a well-researched zonal jet system, we achieved a five-order-of-magnitude speedup in emulating the zonally-averaged flow.
  • This enabled us to efficiently generate large ensembles, facilitating the process of establishing accurate probabilities of rare events, such as transition rates between different long-lived states.
  • Supervisors- Professors Peter Haynes and Rich Kerswell.
  • Python: PyTorch; Julia ; LaTeX - Repository
  • Stochastic Latent Transformer Architecture

    Stochastic Latent Transformer archiecture, consisiting of an Autoencoder structure, with an autoregressive stochastically forced transformer that acts on the latent representations.


Past Projects

Data-Driven Exploration of Mid-Latitude Weather - Master's Thesis - Report

2021

  • Used a Autoencoder to explore whether a beta-plane turbulence model of tropospheric mid-latitude circulation lay on an internal manifold.
  • Looked to evolve the dynamics of the model in the reduced latent space, before exploring the variability of the system due to its stochastic parameterisation scheme.
  • Supervisors- Professors Peter Haynes and Rich Kerswell.
  • Python: Keras, Tensorflow; MATLAB ; LaTeX - Repository

Assessing Temporal Change In The Exposure Of Informal Settlements Through Repeat Satellite Observation - Group Project - Report

2021

  • This project focused on assessing change in the exposure of Caribbean informal settlements over time to natural hazards.
  • Informal settlements were located through segmentation of satellite images using a Random Forest model as well as Deep Learning models.
  • We then used Deep Learning to identify changes to settlement sizes and to quantify vulnerability. For example, following a disaster, change detection algorithms aim to determine the extent of damage suffered (e.g. destroyed, majorly damaged, undamaged).
  • Python - Repository

Quantifying the effecitiveness of natural hazard preventions by using an LSTM to predict rainfall runoff in flood risk mitigation - Group Project

2020

  • Project to investigate the effectiveness of natural flood management interventions undertaken in the town of Shipston-on-Stour during 2017 to 2020 using an LSTM model.
  • Python - Repository

Deep Learning Robustness for Neutrino Event Detection using Adversarial Neural Networks - Bachelor's Thesis - Report

2020

  • Used a Domain-Adversarial Neural Network (DANN) to improve the performance of a Convolutional Neural Network (CNN) to classify neutrino interactions, for the analysis of neutrino oscillations.
  • This method looked to produce a model that is invariant to the differences in statistics between the input data (the labeled Monte Carlo simulations used to train the classifier) and the detector data.
  • Supervisor- Dr Chris Backhouse.
  • Python: Keras, Tensorflow; C++: Root, NOvAsoft; LaTeX - Repository

HPGe Detector Gamma Ray Spectroscopy simulation of nuclear emission and subsequent detector interactions - Group Project Report

2020

Cellular Automata Model to Simulate Traffic Flow's Similarities to Granular Flow - Report

2019

  • Used a Cellular Automata model to simulate motorway traffic flows, in order to compare the similarities to the granular flow, turbulence and choked flow when traffic shockwaves arise.
  • The model consisted of a few rules with the system was able to evolve over time with a stochastic element put in place to represent human decision making and irrationality, and was extended to contain different vehicles with different maximal speeds, blockages such as accidents or road closures to try and model a driverless car system.
  • Supervisor- Professor David Bowler.
  • Python - Repository


Hackathons

Chatbot to translate text in Facebook Messanger Hackathon - Developer Circles from Facebook AI Messanger Hack

2019

  • I was selected to represent Univerity College at the AI for Messenger Hackathon where we created a chatbot that returned the translated text from an image containing text in a different language. Used Node.js for the messenger front end, with Flask connecting to the Pytorch models, which comprised of a CNN to determine the locations of the words, an OCR CNN to recognise the text, and a translation neural network.
  • Image 1; Image 2; Image 3; Image 4; Image 5.
  • Repository.

Providing insight from credit card customer datasets Hackathon - Winning Hackathon Team - University College London Data Science Society Hackathon, hosted by Microsoft and American Express

2019

  • I was part of the winning team, where we produced a solution concluding that that product personalisation for customersubsets could increase credit card growth while assessing potential credit default and delinquency risk. We conducted exploratory analysis through k-means clustering and build decision tree and random forest models using Scikit-Learn and the Azure API in Python.
  • Image 1; Image 2; Image 3.

Adaptive Image Filter Challange - Winning Hackathon Team - Applied Machine Learning Insight Challange at Arm Holdings

2019

  • I was part of the winning team that completed a Python debugging challenge applying an adaptive image filter to a webcamimage using a CNN during an insight into the research being conducted by ARM in the fields of computer vision and natural language processing for mobile devices.
  • Image.

Image Matching Game - Microsoft AI Mini Hackathon - Microsoft Reactor

2018

  • Made calls to Microsoft’s Cognitive Azure API to identify landmarks and animals and run a bot, by altering pre-build code,that played an image matching game against other participants, winning a small prize for my efforts.
  • Image 1; Image 2; Image 3.

Teaching Experience

Supervisions, University of Cambridge

2022-2024 - Part II Computer Science - Deep Neural Networks.
2022-2023 - AI4ER CDT Introduction Classes - Applied Machine Learning.


Organised Workshops

Sep 2023 - Institute of Computing for Climate Science, Cambridge - ICCS Summer School 2023.

  • Organisation Volunteer and speaker - talk title: ’Learning Zonal Jet Dynamics with Stochastic Latent Transformers’.

Mar 2022 - Cambridge Centre for Climate Science - Machine Learning for Climate Science Workshop.

  • Organised with 2 other graduate students a workshop introducing Machine Learning to researchers in the atmospheric sciences.
  • Included with 3 interactive sessions & a hackathon implementing ML on in situ atmospheric data. - In-person attendance reached capacity with 40 attendees, along with 8 virtual attendees.
  • Resources can be found here.

Other Relevant Experience

Pembroke College, University of Cambridge

June 2022 - Nov 2023, Housing Officer, Graduate Parlour.

  • Elected to represent the gradute student body of the college on all matters relating to housing and accomodation.

June 2021 - June 2022, President, Graduate Parlour.

  • Elected to represent the gradute student body of the college, as well as lead the Graduate Parlour committee.

Oct 2020 - June 2021 - Events Officer, Graduate Parlour.

  • Elected to organise events, large and small, that will appeal to all aspects of the college community.
  • This included online events as well as following, and adapting to, Covid protocols to ensure all in-person events are run safely and within guidelines.