Computer Science Researcher Machine Learning for Scientific Problems
I develop machine-learning methods for scientific problems with heterogeneous data, limited labels, and a need for interpretable predictions.
My work studies how models generalise across different populations, regions, and scientific domains, with applications in protein modelling, Earth observation, and earthquake early warning.
Seeking Lecturer positions in Computer Science and MSc/PhD research opportunities. Open to research collaboration.
Research focus
Across my projects, I study how models behave when new data differ from their training data, how their failures can be identified, and how pretrained models can be adapted to specialised scientific problems.
How ML systems behave outside their original training distribution.
Machine learning for scientific data, constraints, and domain questions.
Understanding model decisions and identifying conditions of failure.
Adapting large pretrained models to specialised scientific domains.
Selected research
Ongoing
I study how protein language models can be adapted to thermophile proteins without losing broader biological knowledge. The work compares efficient fine-tuning methods and uses explainability to examine what the adapted model learns.
Ongoing
I develop Sentinel-1 water-segmentation models and test how well they transfer across different regions and environments. I also study how geographic and radar conditions affect model performance, with the goal of identifying where predictions become unreliable.
Published + ongoing
My published study curated 7,318 South Asian earthquake events and compared six models using seven-second, single-station P-wave windows. I am now studying how the model transfers across regions and what early waveform features reveal about shaking and earthquake source properties.
Selected publications
IEEE QPAIN · 2026
N. H. Fahim, A. U. Muhib and M. S. Rahman.
Curated 2,845 protein sequences from nine native Bangladeshi fish species and compared seven architectures for accurate, resource-efficient classification.
IEEE ICCIT · 2025
M. N. Hasan Fahim, M. A. Ullah Muhib, R. A. Tanvir and A. Al Noman.
Curated 7,318 South Asian events and benchmarked six models; the Transformer reached 76.23% accuracy and 81.56% adaptive accuracy at 4.8 ms inference latency.
Teaching & academic interests
I am interested in teaching foundational computer science and machine-learning courses with an emphasis on conceptual understanding, implementation, and problem solving. My primary areas include programming, data structures and algorithms, database systems, software engineering, and machine learning.
Beyond research
I have developed production-oriented web and AI systems using modern backend, database, cloud, computer-vision, and language-model technologies.