Skip to content

The Five Projects

This section is a guide to the five undergraduate-thesis-scale projects provided in this handbook's repository. Each project lives in /projects/0X-name/ and has its own detailed README.md covering background, step-by-step instructions, expected outcomes, and example analysis scripts.

The five projects span the major techniques in this handbook. They are designed to be:

  • Tractable: each fits in 6-10 weeks of part-time work, or 3-5 weeks full-time.
  • Self-contained: the project README has everything you need to start.
  • Open-ended at the top: the basic version is achievable; extensions for ambitious students are sketched.
  • Properly bounded: each has a clear "minimum viable result" so that even a struggling project produces a defensible thesis.

Pick one project. Do it well. Do not try two.

Project 1: Defect Formation Energy in Silicon

Project folder: projects/01-defect-formation-energy/ Duration: ~6 weeks Difficulty: Foundational (you have all the prerequisites) Methods used: DFT only

What you will do

Compute the formation energy of vacancy and substitutional defects in crystalline silicon using DFT. Carry out convergence studies for cutoff energy, k-point grid, and supercell size. Compare your results to published values. Extend (if time allows) to charged defects with image-charge corrections.

Why it is a good thesis project

Silicon is the most-studied semiconductor on Earth. Published values exist for every defect you might compute, so you have many benchmarks. The system is simple enough to converge cleanly on modest hardware. You will learn the full DFT workflow — geometry, k-points, defects, formation energies — without getting bogged down in pathological materials physics.

Expected outcomes

  • Converged formation energies for the silicon monovacancy and one substitutional dopant.
  • Convergence study plots for \(E_\mathrm{cut}\), k-grid, and supercell.
  • Comparison table with two or three published DFT values.
  • (Stretch) Charged-defect formation energies with Freysoldt correction.

Prerequisites

Computational requirements

A modest workstation (16 cores) or a small cluster allocation is sufficient. Total compute: roughly 200-400 CPU-hours including convergence studies.

What can go wrong

Charged defects are surprisingly delicate; the image-charge correction is a subtle topic. If you find yourself stuck, fall back to neutral defects only — they are perfectly thesis-worthy.

Project 2: Melting Point of Copper via MLIP-Driven MD

Project folder: projects/02-melting-point-mlip/ Duration: ~8 weeks Difficulty: Intermediate Methods used: MD + pre-trained MLIP (or classical EAM as fallback)

What you will do

Compute the melting point of copper by molecular dynamics, using either a pre-trained MLIP (CHGNet, MACE-MP-0, or similar foundation model) or a classical EAM potential as a fallback. Use the two-phase (solid-liquid coexistence) method to bracket the melting temperature. Compare to the experimental value (1357 K).

Why it is a good thesis project

Melting points are notoriously hard to compute reliably. Doing this project teaches you finite-temperature MD, the two-phase method (superheating, supercooling, free-energy considerations), proper thermostatting, and how to assess the quality of a force field at a phase transition. The experimental value is known, so you have a hard validation target.

Expected outcomes

  • A clear estimate of the melting temperature with error bars.
  • Equilibration plots, energy histograms, and a phase-coexistence trajectory.
  • Comparison of the MLIP result and the EAM result with experiment.
  • Discussion of why each method succeeds or fails.

Prerequisites

Computational requirements

GPU or moderate CPU resources. With a pre-trained MLIP, total compute is ~100-300 GPU-hours. With EAM, much less; can run on a workstation.

What can go wrong

The two-phase method requires careful setup — you need to construct a solid-liquid interface and watch which side grows. Equilibration is the dominant time cost. Be patient and run multiple long trajectories.

Project 3: High-Throughput Band Gap Screening

Project folder: projects/03-band-gap-screening/ Duration: ~8 weeks Difficulty: Intermediate Methods used: Database query + GNN inference + DFT validation

What you will do

Use the Materials Project database to identify candidate materials with band gaps in a target range (e.g. 1.5-2.0 eV, useful for photovoltaics). Use a pre-trained GNN (e.g. MEGNet, ALIGNN) to predict band gaps for a much larger set of compositions. Validate the top candidates with DFT calculations.

Why it is a good thesis project

This is a complete data-science / high-throughput materials discovery workflow in miniature: a database query, an ML prediction step, an expensive validation step, and an analysis of how well the cheap ML filter performed. You will learn database APIs, GNN inference, and how to handle a high-throughput pipeline.

Expected outcomes

  • A list of 10-20 candidate materials with predicted gaps in your target range.
  • DFT-validated gaps for 5 of these candidates.
  • A precision-recall analysis of the GNN filter.
  • Discussion of the failure modes (where the GNN predicts well, where it fails).

Prerequisites

  • Chapter 10 (GNNs)
  • Chapter 5 (DFT)
  • Some Python and command-line comfort (Materials Project API, Pymatgen)

Computational requirements

GNN inference is cheap (can run on a laptop). DFT validation is the cost: 5 calculations × ~20 CPU-hours each, plus convergence studies. Total ~200 CPU-hours.

What can go wrong

PBE band gaps systematically underestimate experimental gaps by 30-50%. Be clear in your thesis whether you are reporting PBE gaps, gaps with a known scissor correction, or experimentally-comparable gaps from a hybrid functional (which adds cost). Choose your story before launching.

Project 4: Training an MLIP From Scratch on a Novel System

Project folder: projects/04-mlip-from-scratch/ Duration: ~10 weeks Difficulty: Advanced Methods used: DFT (data generation) + MLIP training + MD validation

What you will do

Pick a system that does not yet have a publicly available accurate MLIP (your supervisor will help you identify one — many surfaces, defects, or alloy compositions qualify). Generate DFT training data covering the configurations of interest. Train an MLIP (MACE, NequIP, or simpler choices like SOAP-GAP) on the data. Validate against held-out DFT configurations and against a target property (lattice parameter, phonon spectrum, defect formation energy) where you have a reference.

Why it is a good thesis project

Training an MLIP from scratch is, in a real sense, the defining skill of modern computational materials science. Doing this project gives you experience with data generation, hyper-parameter selection, training-set curation, train/test splits, and out-of-distribution validation. You will leave the project knowing how to build an MLIP for any system you encounter later.

Expected outcomes

  • A trained MLIP with validation MAE quoted and contextualised.
  • A training/validation curve showing convergence with dataset size.
  • Validation of one physical property against DFT.
  • Discussion of where the MLIP can and cannot be trusted.

Prerequisites

  • Chapter 9 (MLIPs)
  • Chapter 11 (active learning, helpful for data selection)
  • Comfort with PyTorch or JAX.

Computational requirements

Substantial. DFT data generation is the dominant cost: maybe 500-2000 CPU-hours. Training: 50-200 GPU-hours. Best done with access to a small GPU cluster.

What can go wrong

The most common failure: the MLIP fits the training data well but fails outside it. Mitigate with proper held-out test sets and, if time permits, an active-learning loop. Be honest in your thesis about the regime of validity.

Project 5: Bayesian Optimisation for Catalyst Composition

Project folder: projects/05-bayesian-catalyst-search/ Duration: ~6 weeks Difficulty: Intermediate Methods used: Bayesian optimisation + GNN oracle (or DFT)

What you will do

Set up a Bayesian-optimisation loop to identify the binary or ternary alloy composition that maximises (or minimises) a catalytic property — for example, hydrogen adsorption energy near an ideal value for the hydrogen evolution reaction. Use a pre-trained GNN as the surrogate "oracle" for the property, or DFT for a smaller search. Compare random-search baselines.

Why it is a good thesis project

Active learning and Bayesian optimisation are central to modern materials discovery. This project gives you hands-on experience with the BO loop, acquisition functions (expected improvement, upper confidence bound), Gaussian processes, and exploration-exploitation trade-offs. The scope is naturally bounded; the loop either converges or it doesn't, and either outcome is writable.

Expected outcomes

  • A BO loop that identifies an optimal composition in fewer steps than random search.
  • Visualisation of the acquisition function and the model uncertainty over the composition space.
  • A comparison of two or three acquisition strategies.
  • Discussion of the limits: where BO works, where it gets stuck.

Prerequisites

  • Chapter 11 (active learning and BO)
  • Some Python comfort (you will use scikit-learn, gpytorch, or BoTorch).

Computational requirements

Modest. The BO loop is light; the oracle dominates the cost. With a pre-trained GNN, the entire project can run on a workstation. With DFT, ~100 CPU-hours.

What can go wrong

The BO loop can converge to a local optimum and miss the global one — this is a feature of the algorithm, not a bug. Discuss it honestly. The other common failure: the oracle is biased, so the "optimum" found is an artefact of the surrogate model.

A comparative table

# Project Methods Duration Compute Difficulty
1 Si defects DFT 6 weeks Low Foundational
2 Cu melting MD + MLIP 8 weeks Moderate (GPU) Intermediate
3 Band gap screening GNN + DFT 8 weeks Moderate Intermediate
4 MLIP from scratch DFT + MLIP training 10 weeks High Advanced
5 BO for catalyst BO + GNN oracle 6 weeks Low Intermediate

How to choose

A few practical considerations.

If you have only used DFT before: Project 1 is the natural extension and gives you the cleanest learning curve.

If you enjoyed MD or are interested in finite-temperature properties: Project 2.

If you are excited about data science and want to learn the high-throughput workflow: Project 3.

If you want the deepest technical project that prepares you for a PhD in computational materials: Project 4.

If you are interested in algorithms and optimisation: Project 5.

If your supervisor has a strong preference: take their preference, unless it conflicts violently with your own. They know more than you do about what will succeed in their group.

How to use the project READMEs

Each project README is structured as follows:

  1. Background — short scientific context.
  2. Question — the specific question you will answer.
  3. Plan — a week-by-week (or stage-by-stage) plan.
  4. Setup — software to install, data to download.
  5. Tasks — concrete tasks, each with expected output.
  6. Validation — how to know you got the right answer.
  7. Extensions — for if you have time spare.
  8. References — starter reading.

Treat the README as a starting document, not a complete recipe. Your thesis will deviate from it as you find unexpected results, hit roadblocks, or discover that a particular step needs more attention than the README anticipated.

That deviation is the project. The README gives you a map; you walk the terrain.

A closing word

You have now read fourteen chapters of methodology and finished with five concrete project options. The remaining ingredient — the one that no handbook can supply — is the patient discipline of seeing a project through from start to finish.

The end of the road is a thesis, written by you, defending claims that you have personally established by computation. That is research.

Exercises — final exercises for the chapter.