Project 4 — Training an MLIP from scratch on a novel system

Research question

Pick a small, well-defined materials system that is not covered well by existing pre-trained MLIPs (a 2D material, a Janus dichalcogenide, a small peptide, a metal–organic-framework fragment, or similar). Generate your own DFT training data, train a MACE potential, run a validation suite, and use the trained MLIP for a short (≥ 100 ps) MD trajectory. Document the failure modes you encounter and the active- learning rounds you needed to fix them.

This is the "deep dive" project: where Project 2 was a clean recipe for a single element, this one asks you to act as a real practitioner making — and recording — every decision.

Why this project

Most MLIP papers report a clean training-set construction, a clean parity plot, and a clean MD trajectory. The reality of training an MLIP from scratch on a system you have not seen before is messier:

• Your first random sampling will miss configurations the MD visits.
• Your first model will explode 5 ps into the production MD.
• Your phonon dispersion will have imaginary modes near the zone boundary.

These are the experiences that teach you what an MLIP is, not what the marketing literature says it is. The project is designed to give you those experiences in a contained, eight-to-ten-week window.

Suggested target systems

Pick one:

1. A Janus dichalcogenide MoSSe monolayer. A 2D semiconductor with a mirror-symmetry-broken sandwich (S–Mo–Se rather than S–Mo–S or Se–Mo–Se). Cell of 6 atoms (1×1×1) or 24 atoms (2×2×1 for sampling). Few existing MLIPs cover this exactly.
2. A 2D borophene polymorph (e.g., β₁₂-borophene). A flat boron sheet with a non-trivial honeycomb topology. Sub-stochiometric boron sheets are not in MP. 24-atom cell.
3. A small peptide (Ala–Pro–Ala) in vacuum. ~30 atoms. Periodic PBC with a large vacuum buffer.
4. A Ti–Zr–Hf MAX-phase precursor (e.g., Ti\(_2\)AlC alloyed with Zr). Substitutional alloy chemistry; MP coverage is sparse for substitutional disorder.

If you have a research lab connection, pick a system that aligns with their work. The methodology is identical.

Expected outcomes

By the end of the project you will deliver:

• A trained MACE potential with documented force MAE on a held-out test set. Target: force MAE within 5 % of the typical force magnitude in the dataset (typically < 80 meV/Å for a covalently bonded system, < 50 meV/Å for a metallic one).
• A validation suite consisting of, at minimum:
• Energy parity plot.
• Force parity plot (separately, per element if more than one).
• MD stability test: a ≥ 100 ps NVE/NVT trajectory at two temperatures, verifying no energy drift and no atom-overlap explosions.
• Radial distribution function (RDF) comparison against the AIMD reference.
• Phonon spectrum at \(\Gamma\) (or, for periodic systems, the full dispersion along a high-symmetry path). The MLIP phonons should agree with DFT phonons to within ≈ 10 % over the spectrum, and no imaginary modes should appear that are not also in the DFT.
• An active-learning log: at least one round in which you identified high-uncertainty configurations during MD, re-labelled with DFT, re-trained, and demonstrated a reduction in the MD failure rate.
• A 4–6 page report describing the system, the choices made, the failures observed, and the final validated potential.

Time estimate

Ten weeks, part-time.

Week	Activity
1	Background reading. Pick the system. Compute DFT lattice parameter and basic properties.
2	Initial data sampling: short AIMD at two temperatures + rattled cells + strained cells. ≈ 300 frames.
3	Train MACE v1. Compute parity.
4	Run first production-like MD with MACE v1. Expect failures. Identify what went wrong.
5	Active-learning round 1: identify high-uncertainty frames (committee of 2 MACEs), re-label with DFT, augment training set.
6	Train MACE v2. Compare validation metrics.
7	Repeat AL if needed. Phonon spectrum check.
8	Final production MD: ≥ 100 ps at two temperatures. RDF check.
9	Write-up.
10	Buffer for cluster outages, model debugging, and revisions.

Compute budget

Variable, depending on system size. Plan for:

• ≈ 1500–3000 CPU-hours for DFT data generation, distributed across AIMD runs + AL relabelling. A 30-atom cell at 100 fs/step on 4 cores is roughly 1 hour per ps.
• ≈ 30 GPU-hours for training MACE (multiple rounds).
• ≈ 200 GPU-hours for production MD if using MACE+LAMMPS at 30 fs per simulated ps on a single GPU.

The 10-week timeline is genuinely tight; if your cluster queues are slow, plan to push the buffer week.

Prerequisites

• Chapter 6 — Running DFT in practice — you must be comfortable running AIMD in QE and extracting forces with full accuracy.
• Chapter 9 — Machine-learning interatomic potentials — the architectural rationale for MACE; the role of body order and cutoff radius; the energy/force-weight trade-off in training.
• Chapter 8 — Statistical mechanics for simulation — the framework for an MD-derived property (RDF, MSD, phonons) as a validation target.

You should have completed Project 2 or have equivalent experience with at least one MLIP training pipeline. This is not a first-MLIP project.

What "covered well" means

A pre-trained MLIP "covers" your system if:

• Its training set includes elements present in your system at similar coordination environments.
• The reported test-set MAE on your system's chemistry is reasonable (typically < 100 meV/Å for a covalently bonded material).
• The MLIP gives stable MD at relevant temperatures.

A foundation model like MACE-MP-0 covers a substantial fraction of crystalline materials. The point of this project is to pick a system where it does not. Verify this by attempting to run MACE-MP-0 on your target — if it gives unreasonable forces or unstable MD, you have the right system. Document this baseline failure as the justification for training from scratch.

Pitfalls flagged up front

1. Picking too big a system. Forty atoms is generous; sixty atoms is challenging; a hundred atoms is a research project, not an undergraduate one. A six-atom cell is fine if the chemistry is the focus.
2. AIMD trajectory length. Two picoseconds at one temperature is not "training data"; it is a single statistically-dependent block. You need multiple short trajectories at different temperatures, plus rattled cells, plus strained cells.
3. Force-MAE vs MD stability. A force MAE of 100 meV/Å on a parity test set does not guarantee MD stability. Some of the worst-error configurations may not be in your test set.
4. Phonon imaginary modes. Small imaginary modes (\(< 5\) cm\(^{-1}\)) near \(\Gamma\) are usually numerical artefacts of the finite- difference Hessian. Larger imaginary modes elsewhere point to a genuinely deficient potential — usually because the relevant strain or twist configuration is absent from training.
5. Cherry-picking the validation set. It is tempting to drop the worst few test-set predictions as "outliers". Don't. Outliers are informative; they are the configurations the model is wrong on.
6. Active-learning hubris. A single AL round usually helps. A tenth AL round on the same system usually doesn't. Know when to stop and report the limitation.

Deliverables checklist

• data/dft/ — extended-XYZ training, val, test datasets, all versions.
• models/v{1,2,...}/ — MACE checkpoints from each iteration.
• validation/ — parity plots, RDF, phonon, MD-stability logs.
• active-learning/round{1,...}/ — uncertainty-flagged frames, DFT relabelling, retraining diff.
• production/md.traj — final long-MD trajectory and analysis.
• report.pdf.