Md. Khairul Islam

Ph.D. Candidate in CS | Generative & Foundation Models for Scientific Data | University of Virginia · Virginia, United States

I design generative and foundation models for scientific data, built to run without a supercomputer and to transfer across instruments and domains. I am a sixth-year Computer Science Ph.D. candidate at the University of Virginia, advised by Professor Judy Fox and co-advised by Professor Geoffrey Fox.

My research follows two threads. In generative and foundation models for science, Cosmo3DFlow (KDD 2026) replaces iterative diffusion sampling with wavelet flow matching, so reconstructing the early universe no longer requires hundreds of denoising steps; OmniSpectra learns one representation of astronomical spectra across instruments, rather than a separate model for each; and serverless inference (IJHPCA 2025) puts such models on demand in the cloud, without dedicated HPC allocations. In explainable time series deep learning, WinTSR (AAAI 2025) recovers the delayed cause-and-effect relationships that point-wise attribution methods miss. Applied to COVID-19 forecasting, this line of work showed which population age groups drove transmission, earning First Place in the 2024 CIC Student Paper Challenge and Third Place at the NSF Student Research Competition, ICDH 2023. I packaged these and 15+ other attribution methods into tslens, a PyTorch library covering 25+ time series architectures.

Before UVA, I spent two years as an iOS software developer at Samsung Research, Bangladesh (2018–2020), after a B.Sc. from Bangladesh University of Engineering and Technology. I have been a teaching assistant for seven graduate and undergraduate courses at UVA (see Teaching). In my spare time I enjoy Chess and Table Tennis.

I am on the job market and available from Spring 2027. If you would like to talk research or explore working together, reach out.

News

Aug 2026Released tslens on PyPI: a PyTorch library unifying 15+ attribution methods across 25+ time series architectures.
Jul 2026Passed my Ph.D. Proposal exam. Title: "Generative and Foundation Models for Scientific Data".
Jun 2026Bronze Medal (top 10%, 279th of 3677 teams) in the Hull Tactical Market Prediction Kaggle competition.
May 2026Cosmo3DFlow accepted at KDD 2026, sampling with an order of magnitude fewer integration steps than diffusion models. Code.
Jan 2026Preprint: OmniSpectra: A Unified Foundation Model for Universal Spectra Representation Learning.
Jan 2026TA for Software Engineering (CS 3240), instructors Prof. Mark Sherriff and Prof. Sarah Elder.
Nov 2025Scalable Cosmic AI Inference using Cloud Serverless Computing accepted at the International Journal of High Performance Computing Applications. Code.
Aug 2025TA for Software Engineering (CS 3240), instructor Prof. Mark Sherriff.
Jul 2025Silver Medal (top 5%, 155th of 3757 teams) in the Jane Street Real-Time Market Data Forecasting Kaggle competition.
May 2025TA for Big Data Systems (DS 5110), instructor Prof. Judy Fox.
Jan 2025TA for Foundations of Data Analysis (CS 3501 / ECE 3502), instructor Prof. Miaomiao Zhang.
Dec 2024WinTSR accepted at the AAAI 2025 Workshop on AI for Time Series Analysis.
Nov 2024Featured in UVA Engineering news, UVA Ph.D. student uncovers COVID-19 transmission patterns.
Nov 2024"Large Language Models for Financial Aid in Financial Time-series Forecasting" accepted at the IEEE BigData 2024 Workshop on LLMs for Finance.
Oct 2024Started astronomy research with NRAO (National Radio Astronomy Observatory) through the NSF–Simons AI Institute.
Sep 2024First Place in the 2024 COVID Information Commons Student Paper Challenge (graduate cohort), organized by Columbia University, for "Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections".
Aug 2024TA for Wireless Sensing for IoT (CS 6501).
May 2024Graduated from UVA with a Master's (en route) degree in Computer Science.
Dec 2023Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections accepted at the AAAI 2024 Workshop on AI for Time Series Analysis. The paper later won First Place in the 2024 COVID Information Commons Student Paper Challenge.
Nov 2023Passed my Ph.D. qualifier exam on interpreting time series models by explicitly accounting for temporal importance.
Oct 2023Temporal Dependencies and Spatio-Temporal Patterns of Time Series Models accepted at the AAAI/SIGAI Doctoral Consortium 2024.
Jul 2023Third Place in the NSF Student Research Competition at the IEEE International Conference on Digital Health, for "Interpreting County-Level COVID-19 Infections using Transformer and Deep Learning Time Series Models".

Selected Publications

Generative & Foundation Models for Science

Cosmo3DFlow: Wavelet Flow Matching for Spatial-to-Spectral Compression in Reconstructing the Early Universe

Md Khairul Islam, Zeyu Xia, Ryan Goudjil, Jialu Wang, Arya Farahi, Judy Fox

KDD 2026

Combines 3D wavelet transforms with flow matching to reconstruct the early universe from cosmological simulations, cutting diffusion-model sampling cost by 46×.

Paper Code

OmniSpectra: A Unified Foundation Model for Universal Spectra Representation Learning

Md Khairul Islam, et al.

Preprint, 2026

A single foundation model for spectra across instruments, handling variable-length inputs at native resolution without resampling.

Paper

Scalable Cosmic AI Inference using Cloud Serverless Computing

Mills Staylor, Amirreza Dolatpour Fathkouhi, Md Khairul Islam, et al.

IJHPCA 2025

A cloud framework combining pretrained models with serverless infrastructure to run deep-learning astronomical inference without dedicated HPC access.

Paper Code

Explainable Time Series Deep Learning

WinTSR: A Windowed Temporal Saliency Rescaling Method for Interpreting Time Series Deep Learning Models

Md Khairul Islam, Judy Fox

AAAI 2025 Workshop (AI4TS)

A windowed saliency method that captures delayed temporal dependencies, outperforming prior interpretation techniques across multiple architectures and datasets.

Paper Code

Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections

Md Khairul Islam, Judy Fox

AAAI 2024 Workshop (AI4TS)

First Place, 2024 COVID Information Commons Student Paper Challenge (graduate cohort)

Benchmarks eight interpretation methods across six transformer models using 3.5M COVID-19 case records to identify which age groups drove transmission.

Paper

Temporal Dependencies and Spatio-Temporal Patterns of Time Series Models

Md Khairul Islam, Judy Fox

AAAI/SIGAI Doctoral Consortium 2024

Doctoral consortium work on explicitly modeling temporal importance to produce more precise explanations of feature interactions in time series models.

Paper Code

See all publications →

Awards

Paper Awards

Sep 2024First Place (graduate cohort) in the 2024 COVID Information Commons Student Paper Challenge, Columbia University, for "Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections".
Jul 2023Third Place in the NSF Student Research Competition, IEEE International Conference on Digital Health (ICDH 2023), for "Interpreting County-Level COVID-19 Infections using Transformer and Deep Learning Time Series Models".

Competition Awards

Jun 2026Bronze Medal (top 10%, 279th of 3677 teams) in Hull Tactical Market Prediction, Kaggle.
Jul 2025Silver Medal (top 5%, 155th of 3757 teams) in Jane Street Real-Time Market Data Forecasting, Kaggle.

See all awards →