Research& Publications
Applied machine learning research in scientific computing and systems security, built on reproducible pipelines and verified baselines.
Research Focus
My research asks practical questions. Can physics residuals flag masquerade attacks in a vehicle network better than purely data-driven models? Can a network trained on the residual of an ODE alone, never shown the answer, recover the solution? Every study runs on a reproducible pipeline with fixed seeds, leakage-proof splits, and metrics chosen to expose weaknesses rather than hide them.
Current Focus
Physics-Informed Neural Networks, Neural ODE Surrogates, ML for Cyber-Physical Systems
Long-term Goal
Graduate research in applied ML & systems
Core Interests
Research Experience
Undergraduate Thesis Researcher (Team Lead)
United International University · Team Paradox • Advisor: Dr. Muhammad Nomani Kabir
Leading FYDP research on the Lorenz-1960 ODE system. The current method is a physics-informed network trained on the ODE residual alone with a hard-constrained initial condition: 11,283 parameters, within 3e-05 of a DOP853 reference, trained in under six CPU minutes. Around it sit a nine-architecture sweep, a windowed PINN study with causal residual weighting for longer horizons, and the ICEEICT 2027 manuscript. FYDP-II defended.
Independent Researcher
PINN-CPS-Robo · CAN Masquerade Attack Detection • Advisor: Self-directed
Studying whether physics-residual signals detect CAN masquerade attacks on the ROAD dataset better than purely data-driven baselines, with attention to attack attribution.
Researcher
DT-NB Hybrid Study • Advisor: Self-directed
Compared Naive Bayes instance filtering and decision-tree attribute weighting, separately and in both orders, with all supervised steps refit inside each training fold across ten seeds. Manuscript written; not yet submitted to a conference.
Independent Researcher
AI-Generated Image Forensics • Advisor: Self-directed
A pre-registered feature-level study of pixel-pattern, frequency, and combined feature sets for detecting AI-generated images under matched JPEG compression and unseen-generator shift.
Solo Reviewer
PINN Literature Review Dossier • Advisor: Self-directed
A structured survey of physics-informed neural networks covering 2021-2026 plus pre-2021 anchors, formatted for Springer's Artificial Intelligence Review.
Undergraduate Teaching Assistant
United International University • Advisor: CSE Department Faculty
Supporting students in DBMS, DSA-I, ICS, and SWE lab courses. I run lab sessions and tutorials, review code, and grade assignments.
Publications
NB Instance Filtering and Tree-Based Attribute Weighting: A Hybrid Study on Noisy Tabular Data
Manuscript in PreparationIkramul Hasan Moral*
Self-directed study · Manuscript written, not yet submitted
A conference-ready study comparing Naive Bayes instance filtering and decision-tree attribute selection/weighting, separately and in both orders, with all supervised steps refit inside each training fold and scored on held-out rows across ten seeds. The manuscript is complete but has not been submitted to a conference.
Lorenz-1960 PINN: Solving the ODE System from the Equations Alone
Defended · Submitted to ICEEICT 2027Ikramul Hasan Moral*, Md. Abu Bakar*, Samiur Rahman Omlan*, Fariha Islam*, Md. Touhidul Islam*
UIU Final Year Design Project · Supervisor: Dr. Muhammad Nomani Kabir
A physics-informed network with a hard-constrained initial condition, trained on the ODE residual alone with no data, lands within 3e-05 of a DOP853 reference in under six CPU minutes. A nine-architecture sweep and a windowed causal-weighting study for longer horizons sit alongside the run of record. Sixteen paper runs plus sixteen windowed models are archived with SHA-256 verification, and the manuscript is submitted to ICEEICT 2027.
ANN Modeling of Hybrid Nanofluid Boundary Layer Flow
Active · Extending to New EquationsIkramul Hasan Moral*
Independent scientific ML project
A nine-layer neural network trained with Levenberg-Marquardt optimization to model hybrid nanofluid flow and heat transfer over a stretching sheet, validated against numerical solutions with MSE, RMSE, and R-squared metrics. The boundary-layer direction is active again, now aimed at new equation systems.
Talks & Teaching
Research Topics
Active Research Areas
PINNs on New Equations
Extending the defended Lorenz-1960 thesis: windowed PINNs with causal residual weighting for longer horizons, and boundary-layer problems with new equation systems.
Physics-Residual Attack Detection
Studying physics-residual signals for CAN masquerade detection on the ROAD dataset, and PINN-based shape optimization for small robots (PhORM).
AI-Generated Media Forensics
A pre-registered feature-level study of which simple feature sets detect AI-generated images under matched compression and unseen generators.
Research Data Collection
Built a live web instrument for collecting human perception data on memes for an ongoing research study.