Undergraduate Literature Review Guide

Example Topic: Reconstruction and Forecast of Spatiotemporal Data in Physics: Multi-Scale Challenges

---

🎯 Objective

You will conduct a structured literature review to support the introduction section of a research publication.
Your goal is to explore how scientists and engineers reconstruct and forecast multi-scale spatiotemporal physical data, and to compare the state of the art with and without machine learning (ML) approaches.

Your final review will:

1. Explain why the problem (e.g., reconstructing and compressing multi-scale wavefields) is important, difficult, and impactful if it were resolved..
2. Summarize traditional (non-ML), ML-based approaches, and whatever is the state of the art.
3. Highlight what advanced architectures (e.g., SHRED, CNNs, Transformers, PINNs) or curated data sets (e.g., balanced and diverse, data volumes; e.g., The Well) can do — and where their limitations and opportunities.

---

🧭 Step-by-Step Instructions

1. Define the Problem

• Write a short paragraph explaining what the problem (e.g., reconstruction and forecasting) mean in the context of physical systems (e.g., wavefields, turbulence, seismic data, or climate), and ensure to write a clear statement for why it will be impactful if solved for society, for engineering, etc.
• Explain what aspects of the problems (e.g., multi-scale problems) are challenging — mention scale separation, nonlinearity, and high-dimensionality.
• State that your goal is to understand how data-driven and physics-based approaches address these challenges.

---

2. Structure of the Literature Review

Your report should have the following sections:

A. Introduction and Motivation

• Define the general problem and societal impact.
• Describe why the problem (e.g., reconstructing and forecasting spatiotemporal) matters in geophysics.
• Identify multi-scale examples (e.g., wave propagation, turbulence, geophysical imaging).

B. Traditional / Physics-Based Approaches

• Discuss non-ML methods (e.g., PDE solvers, spectral methods, compressed sensing, and Kalman filters).
• Note their successes and limitations in high-dimensional or multi-scale systems.

C. Machine Learning-Based Approaches

• Summarize ML frameworks (CNNs, RNNs, Transformers, autoencoders, GANs, SHRED architectures).
• Discuss how they model spatial and temporal correlations.
• Highlight examples and benchmarks from physics or geosciences.

D. Limitations and Open Challenges

• Describe what remains difficult: generalization, interpretability, physical consistency, scalability.
• Identify hybrid methods (e.g., physics-informed networks) and what problems they solve.

E. Synthesis and Research Gaps

• Summarize 2–3 key research gaps that motivate further work.

---

🔎 How to Find and Select Literature

1. Use Multiple Search Platforms

🔹 Google Scholar

• Start with Google Scholar.
• Use targeted search terms such as:
  • spatiotemporal reconstruction physics
  • multi-scale forecasting physical systems
  • wavefield reconstruction machine learning
  • physics-informed neural networks multiscale
  • SHRED architecture geophysical data
• Sort by recent papers (e.g., since 2018) to find the latest work.
• Read the introductions of these papers to find more papers to read or use the “Cited by” feature to discover related or follow-up studies.

🔹 ASTA (Allen Institute for AI)

• Explore the ASTA search engine.
• It provides semantic access to scientific papers and can help identify connected ideas across disciplines.
• Use it to find papers related to physics-based ML, wavefield modeling, and spatiotemporal forecasting.

🔹 Researchers’ Websites

• Visit websites of researchers who publish in this area (e.g., in geophysics, applied math, AI for physics).
• Look for:
  • Preprints or open-access PDFs not yet indexed by databases.
  • Project pages or GitHub links that provide code, figures, and datasets.
• Identify multiple research groups from different institutions and countries to ensure your review reflects a broad perspective.

---

2. Read Strategically

• Follow the citation trail:
  When reading a paper, examine what it cites in its introduction. These cited works often explain foundational ideas or widely accepted methods — read them next.

• Diversify your sources:
  Cite papers from at least 3–4 different research groups in different institutions or countries to ensure breadth and representation of the global research landscape.

• Assess quality and relevance:
  Prioritize peer-reviewed papers but include influential arXiv preprints when relevant to recent machine learning advances.

• Read efficiently:
  Start with abstracts, introductions, and figures to grasp the method and findings before reading the entire paper.

---

3. Keep Notes Systematically

Create a table or spreadsheet like this to track your findings:

---

✍️ Writing Guidelines

• Length: 4–6 pages (≈1500–2500 words)
• Include:
  • Clear introduction and motivation
  • Well-organized review sections
  • At least one figure or conceptual diagram (with citation)
  • Consistent reference style (APA, IEEE, or similar)
• Use formal academic writing — avoid bullet lists in the final text.

---

---