What Research Methodology Taught Me: A Post‑Mortem of My Most Transformative Course

The Uncomfortable Truth About “Research”

Before this semester, I thought “research” meant reading a few papers, running some code, and writing a report. I was wrong. Research Methodology is not a box‑checking exercise. It is the scaffolding that turns a pile of ideas into a credible, reproducible, and impactful contribution. Over four months, I learned that a well‑constructed abstract, a reproducible SLR, and a defensible problem statement are worth more than a hundred mediocre experiments.

This blog is my raw, unfiltered reflection — what I did right, where I failed, how I prepared, and what every graduate student should steal from my experience. I’ve packed it with the jargon, frameworks, and mental models that finally made research click for me.

1. The Course Architecture — From Chaos to PRISMA

The professor structured the semester around three pillars: Research writing (abstracts & problem statements), Systematic Literature Review (SLR), and a hands‑on individual project (lightweight deep learning on an ESP32). Assignments were not busywork — every task fed directly into the final exam and our thesis planning.

What I did right

• I treated every lecture transcript as a goldmine. The professor’s off‑hand remarks (e.g., “SLR is like a database query — same input, same output”) became my exam anchors.
• I built cheat sheets for PRISMA, Boolean query dimensions, and the FAST abstract model. Those sheets saved me during the 3‑hour final during revisions.

Where I failed (and you shouldn’t)

• Underestimating the literature review table. I initially skimmed papers and collected vague notes. Only after failing my first draft did I realise that each column (dataset, technique, limitation) must directly answer a research question. I lost two weeks re‑extracting data.
• Ignoring grey literature early. I stuck to peer‑reviewed journals, but my topic (edge AI for AQI) had cutting‑edge pre‑prints on arXiv that completely changed my understanding of quantisation.

2. Abstract & Problem Statement — The FAST Model That Changed My Writing

The professor drilled the FAST acronym into us:

| F | Find the problem & its importance | | A | Approach / claim | | S | Sketch methodology | | T | Tell results (numbers!) |

My mistake: In my first abstract, I wrote “superior performance” without numbers. The professor circled it in red: “Superior to what? By how much?” I learned that without RMSE, accuracy, or latency numbers, your abstract is mere opinion.

My winning move: I rewrote my project abstract using FAST. Now it reads:

“Air pollution (PM₂.₅) causes millions of premature deaths annually, yet most deep learning models rely on cloud servers, causing latency and privacy risks (F). This paper proposes a hybrid 1D CNN‑LSTM quantised to 8‑bit integers and deployed on an ESP32 microcontroller (A). We preprocess the Beijing PM2.5 dataset using a 24‑hour sliding window, train the model, then apply post‑training int8 quantisation (S). The quantised model achieves RMSE 9.04 and MAE 6.45 — only 4.5% higher than float32 — while inference takes 45 ms and uses 65 kB RAM (T).”

Advice for you: Before you write a single line of your thesis, practise FAST on three random papers. Time yourself. You will never again write a vague abstract.

3. Systematic Literature Review (SLR) — PRISMA, Snowballing, and the Joy of Reproducibility

The SLR was the course’s crown jewel. We followed the PRISMA 2020 framework. My group reviewed 100 papers on deep learning for weather‑driven AQI prediction. Here is what stuck:

The PRISMA flow (memorise this)

1. Identification — 2,044 records from Scopus + Google Scholar.
2. Screening — 488 duplicates removed → 1,472 screened → 1,349 excluded at title/abstract.
3. Eligibility — 123 full‑text articles + 51 from snowballing → 174 assessed → 71 excluded.
4. Included — 103 studies for synthesis.

Snowballing — the secret sauce

I initially missed half the relevant papers because my query was too narrow. Backward snowballing (checking reference lists) and forward snowballing (papers that cite your set) unearthed 51 extra studies. Lesson: If your search returns <20 papers after screening, you have missed synonyms — go snowball.

Inclusion/exclusion criteria — where rigour lives

We defined four INCL and four EXCL criteria. Example:

• INCL2 — study uses at least one deep learning technique.
• EXCL2 — study focuses on optimisation for LLMs (reverse direction).

My failure: I initially forgot to include a “language” criterion (only English). That introduced language bias — a validity threat. My teammate caught it. Fix: Always list language, publication year, and study type in your criteria.

Quality assessment — separating wheat from chaff

We scored each paper on internal validity, external validity, reliability, and objectivity. A paper that ran an LSTM only once (no repeated runs) lost points — that’s poor reliability. Takeaway: Quality assessment is not gatekeeping; it’s the only way to trust your synthesis.

4. Research Questions — From “Is X good?” to “How much?”

The professor hated binary, vague RQs. “Is deep learning good for AQI?” → F. “How does a quantised LSTM compare to a GRU in terms of RMSE and inference time on the Beijing dataset?” → A.

My corrected RQs (from my SLR):

1. Which deep learning architectures (LSTM, CNN, Transformer, GNN) are most frequently used for weather‑driven AQI prediction, and what are their reported RMSE values?
2. What publicly available datasets are commonly used, and what are their spatial/temporal resolutions?

Pro tip: Write your RQs before you search. If you can answer them with a yes/no, rewrite.

5. The Individual Project — Lightweight Deep Learning on ESP32

This was the hands‑on validation of everything we learned. I built a hybrid CNN‑LSTM model, quantised it to int8, and deployed it on an ESP32 — a $5 microcontroller with 320 kB RAM.

What I did right

• Architecture: 32 filters (Conv1D) → MaxPool → LSTM(32) → Dropout → Dense(1). ≈5,300 parameters — tiny enough for embedded inference.
• Quantisation: Post‑training int8 reduced file size from 98 kB to 17 kB (5.8× compression) with only 4.5% RMSE increase.
• Evaluation metrics: I reported not only RMSE/MAE but also inference time (45 ms), RAM usage (65 kB), and flash footprint (1.2 MB). Those edge‑specific metrics made the paper publishable.

What I did wrong

• No real sensor data. I used simulated input. A real PM2.5 sensor (e.g., PMS5003) would have added immense credibility.
• Single‑city dataset (Beijing). External validity is low. Future work must include Delhi, Lahore, and London.

Advice for future students

• Start with a tiny model. Do not begin with a 10‑layer transformer. A shallow CNN‑LSTM trained on a Colab GPU converges in minutes and is much easier to quantise.
• Document every hyperparameter. I lost reproducibility because I forgot to record the early stopping patience value. Use a table in your logbook.

6. Visualisations — Making Your Data Talk

I learned that a picture is worth a thousand words, but a box plot is worth ten thousand. For my SLR, I used:

• Bar chart — frequency of architectures per year.
• Box plot — distribution of RMSE across studies.
• PRISMA Sankey diagram — flow of papers.

Exam tip: If asked to suggest a visualisation for “Which transformer variant is most popular for edge AQI?”, say horizontal bar chart — it lets you read variant names easily.

7. The Final Exam — What the Previous Batch Faced

A senior told me the exam had two heavy‑weight questions:

1. Write a PhD research proposal (on any topic).
2. Write a research paper based on your semester project.

How I prepared for the PhD proposal

I memorised a 12‑section template (Title, Abstract, Introduction, Problem Statement, RQs, Aim/Objectives, Preliminary LR, Methodology, Expected Outcomes, Feasibility, Ethics, References). I practised on a topic I knew: “Federated Learning for Privacy‑Preserving AQI Prediction Across Multiple Cities.” In the exam, I adapted it on the fly.

Key for you: Keep your proposed methodology concrete — mention datasets, algorithms, and evaluation metrics. A vague “we will use deep learning” will cost you marks.

How I prepared the research paper

I literally re‑wrote my individual project as a compact IMRaD paper: Introduction, Methods, Results, Discussion. I forced myself to include numerical tables and figure captions. When the exam asked for it, I wrote it in 45 minutes.

Mistake to avoid: Do not copy‑paste your assignment verbatim. The exam expects a condensed, stand‑alone paper — shorter than the full report but containing all key sections.

8. Where I Failed and What I Wish I Had Done Differently

Failure

Not using a reference manager from day 1

Skipping the “first pass” reading

Ignoring the discussion section in papers

Over‑engineering the model

Not collaborating on the SLR extraction

What I learned

I wasted hours reformatting BibTeX. Start with Zotero or Mendeley, and export in IEEE style directly.

I read every paper from start to end. The three‑pass method (abstract → skim → deep read) would have saved 30+ hours.

The limitations and future work paragraphs are gold for finding research gaps. I only realised this in week 10.

My first model had 50 layers. After failing to quantise it, I realised simplicity is a feature, not a bug.

I tried to extract data from 50 papers alone. Division of labour (each team member handles 20 papers with a shared spreadsheet) is the only sane way.

9. How You Can Adapt My Learnings — A Cheat Sheet for Future Students

Before the semester starts

• Take a free DataCamp course on AI ethics and prompt engineering (your professor will likely assign them).
• Install Zotero and learn to import PDFs from Google Scholar.

During week 1–4 (Abstract & Problem Statement)

• Memorise FAST and the three components of a problem statement.
• Practise evaluating abstracts: read one abstract per day and rate it Excellent/Good/Average/Poor using FAST. Share your rating with a peer.

During week 5–10 (SLR)

• Define your RQs before you write your search query.
• Create a 4‑dimension Boolean query (Topic, Methodology, Application, Platform). Test it on Google Scholar; if results <100, add synonyms.
• Draw the PRISMA flow diagram as soon as you have your numbers — it forces you to be honest about exclusions.
• Build an Excel extraction sheet with at least 15 columns (title, authors, year, dataset, method, metrics, limitations, etc.). Have two team members extract the same 5 papers to calibrate.

During the individual project

• Start with a tiny, quantisable model. A 2‑layer LSTM is fine. Do not chase SOTA on day one.
• Measure everything — training time, inference time, RAM, flash, power (if possible).
• Write your final paper incrementally — abstract last, methods first.

For the final exam

• Memorise the PhD proposal template (at least the 8 core sections). Practise writing a proposal in 60 minutes.
• Re‑write your project as a 4‑page paper (IMRaD) and time yourself.
• Know your numbers (RMSE 9.04, 45 ms, 65 kB, 5,300 parameters) — they are your ammunition.

10. The Single Best Thing About This Course

It is not the content. It is the mindset shift. Before RM, I thought research was about being “smart”. Now I know it is about being systematic, transparent, and humble. The professor repeated: “If you cannot reproduce it, you did not do it.” That phrase will guide me.

To the next batch: Do not fear the 3‑hour exam. Fear the 300‑hour journey that makes the exam trivial. Use this blog as your roadmap. If I, with all my failures, can finish, so can you.

ResearchMethodology #SystematicLiteratureReview #PRISMA #EdgeAI #TinyML #AQIPrediction #ESP32 #DeepLearning #QualitativeVsQuantitative #PhDProposal #ResearchQuestions #AcademicWriting #FASTNUCES #GradSchool #ThesisHacks #SLR #Snowballing #InternalValidity #ExternalValidity #GitHubForResearch