Research — Md Nasiat Hasan Fahim

Research vision

Reliability beyond the training distribution

Scientific machine-learning models often encounter data that differ from their training data. Sensors change, geographic regions differ, biological conditions vary, and labelled data are often limited.

My research studies how these differences affect model performance and how failures can be detected or reduced. I am particularly interested in generalisation, explainability, uncertainty, and efficient model adaptation.

I explore these questions using protein sequences, satellite radar, and seismic waveforms.

Current research 01

Scientific foundation models

Ongoing

Manuscript submitted
2026

Adapting protein language models to thermophile sequence space

Research question: How well can a general protein language model represent proteins adapted to extreme environments?
Challenge: Protein language models are trained on proteins from organisms living under very different environmental conditions. Apparent improvements can also come from near-duplicate sequences rather than genuine adaptation.
Approach: I adapt ProGen2 using LoRA-based supervised fine-tuning, DPO, and KTO under matched compute budgets and homology-controlled evaluation.
Contribution: I compare likelihood-based and preference-based adaptation methods under the same compute and evaluation conditions to understand their different strengths and trade-offs.
Current direction: I use attribution methods to study which residue, charge, and sequence patterns the adapted models associate with thermophile biology.
Resources: Submitted manuscript details

ProGen2LoRA / PEFTDPOKTOHomology-controlled evaluationXAI

Current research 02

Geospatial and Earth observation ML

Ongoing

Sentinel-1
2025 –

Cross-regional water segmentation from Sentinel-1 SAR

Research question: How far can a water-segmentation model move across regions and biomes before its predictions become unreliable?
Challenge: SAR backscatter changes with wind, incidence angle, vegetation, terrain, speckle, shadow, and layover. Labels are limited, and geographic validation can hide local failure modes.
Approach: I train and compare cross-regional segmentation models, evaluate geographic and biome shift systematically, and inspect model behaviour with explainability methods.
Contribution: I focus on identifying the geographic conditions under which segmentation performance degrades, rather than reporting only average accuracy.
Current direction: I am developing a controlled evaluation setup for testing cross-region performance and using explainability to understand where Sentinel-1-only models fail.
Resources: Code and data are not yet publicly released.

Sentinel-1Semantic segmentationDomain shiftGeoAIXAI

Current research 03

Earthquake early warning

Published + ongoing

IEEE ICCIT 2025
Extension in progress

Early inference from initial P-waves

Research question: What earthquake information can be inferred reliably from the first seven seconds of one station's P-wave, and how well does that inference transfer across regions?
Challenge: An early-warning system must make a decision before the damaging S-wave arrives. The data are imbalanced, stations and instruments vary, and a regional model can learn local artefacts.
Approach: I curated 7,318 South Asian events and compared logistic regression, gradient boosting, CNN, residual CNN, BiLSTM, and Transformer models under a shared protocol.
Contribution: The study established a regional evaluation dataset and a six-model comparison. The Transformer reached 76.23% accuracy, 81.56% adaptive accuracy, and 4.8 ms inference latency.
Current direction: I am extending the work to evaluate cross-region transfer and to study what early P-wave features reveal about shaking, radiated energy, and source properties. Explainability methods are used to examine which waveform features drive the predictions.
Resources: Peer-reviewed paper · Publication details

SeismologyTransformersCross-region transferReal-time inferenceXAI

Earlier and supporting research

Applied biological and social data

Published · IEEE QPAIN 2026

Protein-based fish species identification

I curated 2,845 protein sequences from nine native Bangladeshi fish species and compared seven architectures. A motif-aware CNN-Transformer with type-aware positional encoding achieved 79.80% accuracy while remaining suitable for CPU inference.

DOI Details

Research assistantship · 2025–2026

Spatiotemporal crime dynamics in Sylhet

Under Dr. Ahsan Habib at SUST, I analysed newspaper and police data using K-means clustering, time-series forecasting, recidivism analysis, and 3D visualisation of crime clusters and victim-offender relationships.

Manuscript details

Future direction

Where I want to take this

I want to develop models that can recognise when they are outside their training conditions, estimate how reliable their predictions are, and adapt efficiently to new scientific domains.

This direction connects my interests in generalisation, uncertainty, explainability, and scientific foundation models.

What breaks when a model leaves home