ANRF AISEHack 2.0 — Polymer Property Prediction

Predict seven physical properties of a polymer — thermal, electronic, optical and dielectric — from its repeat unit written as a PSMILES string. Scored on the unweighted mean R² across all seven, which means a property with 221 training labels counts exactly as much as one with 4,143.

Our final system is a five-model ensemble combined by a per-property cross-fitted Ridge and then corrected by an explicit, affine-fitted physics stage. The central modelling idea is treating the dielectric constant not as a fitted correlation but as an exact decomposition, ε = n² + ε_ionic, licensed by how the reference DFPT calculations are defined.

📖 Read the full journey — how we got here, what broke, and the two discoveries that decided it.

• The stack beats its best member by +0.0151 — that is what decorrelation buys.
• Physics adds a further +0.0105, concentrated exactly where the relations reach: ei +0.0250, eps +0.0244, nc +0.0168, eea +0.0070.

For scale: at our mid-competition leaderboard snapshot, first place on the public board stood at 0.917 — we finished the finale above it, at 0.924.

That +0.0006 gap is the number we are proudest of, and it is not luck. We started Round 3 with an inherited pipeline whose OOF read 0.903 and whose leaderboard read 0.883 — a −0.020 lie. Closing it took a deliberate leak hunt (four separate leaks, all of one shape) and the removal of early stopping everywhere: a model that picks its iteration count by watching the rows it is then scored on reports a number it cannot reproduce out of sample.

ANRF AISEHack 2.0 — a national AI-for-Science hackathon hosted on Kaggle, with the Grand Finale at IIT Madras. Three elimination rounds, each a private community competition scored on mean R² across polymer properties.

flowchart LR
    R1["<b>ROUND 1</b><br/>2 properties<br/>145 teams<br/><br/><b>30th</b>"]
    R2["<b>ROUND 2</b><br/>7 properties<br/>76 teams<br/><br/><b>6th</b>"]
    R3["<b>ROUND 3 · FINALE</b><br/>7 properties<br/>+ 3 judged themes<br/>10 finalists<br/><br/><b>3rd of 145</b> 🥉"]

    R1 -->|qualified| R2 -->|qualified| R3

    style R1 fill:#f5f5f5,stroke:#999,color:#000
    style R2 fill:#e0f0ef,stroke:#00807F,color:#000
    style R3 fill:#fdf0d9,stroke:#B8790B,stroke-width:3px,color:#000

145 teams entered the hackathon; 10 survived to the finale, where we placed 3rd.

Round 3 deliberately moved the goal past raw accuracy. From the competition overview: "the challenge moves beyond raw accuracy to robustness and trust — two qualities that separate a research prototype from a model that materials scientists can actually rely on." It also added a 5.97 M-molecule auxiliary SMILES corpus and a hard rule set — one notebook execution does everything, no external data, no pretrained weights, no uploaded artifacts, seeds fixed and printed.

This repository contains our Round 3 work. Rounds 1 and 2 are kept locally but are not published here.

Score = mean( R²_tg , R²_egc , R²_egb , R²_eps , R²_nc , R²_ei , R²_eea )     ← UNWEIGHTED

tg has 4,143 training rows and eea has 221 — and they count exactly the same. The five smallest properties are 5/7 of the score sitting on under 4% of the rows.

So: model per target, or standardise per target before any shared loss. Never optimise one pooled loss on raw targets — tg reaches 495 and nc reaches 2.76, so tg would swallow the gradient and every other property would quietly collapse to its mean.

flowchart TD
    IN["<b>12,345 PSMILES</b><br/>train 7,405 · test 4,940<br/>every repeat unit has exactly two * points"]
    CANON["<b>CANONICALISE AT INGEST</b> (RDKit)<br/>10,605 raw strings → 8,990 distinct molecules<br/><i>makes every later stage a function of the MOLECULE</i>"]
    FEAT["<b>FEATURISE → 3,037 columns</b><br/>Morgan r2/r3 · atom-pair · torsion · MACCS<br/>RDKit descriptors · SMARTS groups · intensive twins<br/>+ leak-guarded co-observed partner block"]
    ARCH["<b>ROUND-2 ARCHIVE</b><br/><i>host-sanctioned</i><br/>6,165 tg/egc labels<br/>① +2,446 training rows<br/>② partner labels 15.3%→38.1%<br/>③ measured override"]

    GBDT["LightGBM · XGBoost · CatBoost<br/>70 models each"]
    NN["Multi-task NN<br/>6 seeds × 10 folds"]
    GNN["<b>Periodic GNN</b><br/>4 seeds × 110 epochs<br/><i>best single model 0.8978</i>"]

    STACK["<b>PER-PROPERTY CROSS-FITTED RIDGE STACK</b><br/>five columns → one Ridge per target<br/>OOF 0.9129"]
    PHYS["<b>TWO-PASS PHYSICS BLEND</b> — every relation AFFINE-FITTED<br/>ei = egc + eea · eea = ei − egc · egb ≈ egc<br/><b>eps = nc² + ε_ionic</b> ← a DFPT identity, not a regression<br/>OOF 0.9233"]
    POST["Partner regression → clip to observed range"]
    OUT["<b>ARCHIVE OVERRIDE → submission.csv</b><br/>measured truth on 2,450 of 4,940 test rows<br/>applied AFTER clipping<br/><b>Public LB 0.924</b>"]

    IN --> CANON --> FEAT
    ARCH -.-> FEAT
    FEAT --> GBDT & NN & GNN
    GBDT & NN & GNN --> STACK --> PHYS --> POST --> OUT
    ARCH -.->|measured values never perturbed| OUT

    style IN fill:#ffffff,stroke:#999,color:#000
    style CANON fill:#e0f0ef,stroke:#00807F,color:#000
    style FEAT fill:#ffffff,stroke:#999,color:#000
    style ARCH fill:#fdf0d9,stroke:#B8790B,stroke-width:2px,color:#000
    style GBDT fill:#f5f5f5,stroke:#999,color:#000
    style NN fill:#f5f5f5,stroke:#999,color:#000
    style GNN fill:#e0f0ef,stroke:#00807F,stroke-width:2px,color:#000
    style STACK fill:#e0f0ef,stroke:#00807F,color:#000
    style PHYS fill:#fdf0d9,stroke:#B8790B,stroke-width:2px,color:#000
    style POST fill:#ffffff,stroke:#999,color:#000
    style OUT fill:#e0f0ef,stroke:#00807F,stroke-width:3px,color:#000

The full annotated pipeline, with every number taken from the saved outputs of the submitted run:

Five base models, fold-averaged, with no early stopping anywhere. Each model first averages its own fold replicas into one vector; those five vectors enter a per-property Ridge as five columns. There is no plain averaging across models anywhere, and the combiner is cross-fitted so its weights never see the rows they are scored on.

Per-property KFold, not polymer-grouped folds. This is the opposite of the usual advice, and it is measured:

Holding out a whole polymer removes the other properties measured on it — but at test time those properties are in train.csv for nearly every DFT-block row. Grouped CV would have measured a harder problem than the one being scored. (Within a single property there are essentially zero duplicate canonical polymers, so plain KFold is already leakage-safe.)

Physics applied after stacking, not as a feature. A tree splits one axis at a time and cannot cheaply represent a sum of two columns, so ei = egc + eea is wasted as one feature among 3,037. As an explicit post-stack blend it is worth +0.0250 on ei.

Three of our five relations were nearly exact when applied raw. Two were not:

That collapse to 0.336 is not noise — it is physics. Maxwell's relation holds at optical frequency, but the dataset tabulates the static dielectric constant. A DFPT calculation accounts for exactly two contributions and nothing else — an electronic part and an ionic part — and those are stored as separate fields in the dataset of record. Since the electronic part is n², the shortfall is exactly the ionic term:

$$\varepsilon_{ionic} = \varepsilon - n^2$$

An identity, not a regression — so eps − nc² is not a residual to be absorbed by an intercept. It is a physical quantity with its own structure–property relationship, and it can be modelled.

Two independent checks that this was real rather than convenient:

• eps − nc² is positive on all 134 co-observed molecules (mean +0.767, min +0.024). A generic empirical residual would cross zero; an ionic contribution cannot be negative.
• Our own blind affine fit had already converged on a = 1.040, b = 0.615 — a slope of one and an intercept equal to the mean ionic term. The regression was rediscovering the decomposition on its own, with the ionic part pinned to a single constant.

Naming it correctly makes the constant predictable:

The lesson: reading where the data came from paid more than any amount of further tuning. The same fit, correctly named, went from a constant to a model.

The official baseline notebook shipped with the competition fetches two Google Drive files. Their -O names are swapped: the file being written to test.csv is in fact a labelled table of 6,171 tg/egc values from the Round-2 release.

We asked the hosts directly whether it was in scope, and they confirmed it — their own baseline downloads it. The loader (Round 3/src/archive.py) then validates the file rather than trusting the filename, declares itself in the compliance audit, and falls back to running archive-free if it is unavailable.

Measured against Round 3: 3,719 of its labels were already in train.csv and 2,450 correspond to test rows, with values agreeing on 3,717 of 3,719 overlaps — the same measurements, not a second dataset. Used three ways:

1. +2,446 training rows — 7,405 → 9,851 (+33%). The boosters are per-property so they only see more tg/egc, but the NN and GNN share one trunk across seven heads, so a larger abundant-target block improves the representation the five scarce targets ride on.
2. Partner labels — true_egc test coverage 15.3% → 38.1%.
3. A measured override on 2,450 of 4,940 test rows (49.6%), applied after clipping so a real measurement is never perturbed downstream.

A repeat unit is not a molecule with two dangling stubs: the two * points are the same bond to the neighbouring unit. So we drop both dummies, bond their neighbours, and flag the wrap-around edge — the network then encodes an infinite chain.

It became our best single model (0.8978) and the most complementary one: worse than the boosters on egc, far better on eea (0.9264 vs 0.8758) and ei (0.8777 vs 0.8367). It is the only member not built on the shared 3,037-column matrix, so its errors are structurally decorrelated. Adding it took the stack from 0.8713 to 0.9129.

And the readout is an invariance decision, not a hyperparameter. A sum readout scales with repeat-unit count; a mean does not. Measured on an untrained network — a property of the architecture, not of any fit:

Four orders of magnitude — which is why we never use a sum readout.

Round 3 scored robustness, explainability and invariance alongside accuracy. We treated all three as machine-checked deliverables the notebook proves at runtime, not claims on a slide.

Canonicalising at ingest is the mechanism: it makes every downstream stage a deterministic function of the molecule, and canonical(rewrite(s)) == canonical(s) for any valid rewriting. It was never hypothetical — 67.6% of the supplied SMILES strings are non-canonical.

Computed inside LightGBM via pred_contrib=True — the same algorithm the shap package implements, deliberately with no external dependency that could fail to install mid-run.

• tg — intensive twins carry 25.4% of attribution, ring counts lead. Chain stiffness dominates, exactly as free-volume theory predicts. (The invariance design choice turned out to be the biggest accuracy lever on the biggest property.)
• egc — rd_FractionCSP3 tops the list: an sp³-rich backbone localises electrons and widens the gap; conjugation narrows it.
• egb — the fitted partner combination is 26.0% and true_egc 10.3%. The relation egb ≈ egc appears in the attribution rather than being asserted by us.
• nc / eps — molar-refractivity surface descriptors appear for both, reproducing refractive index as electronic polarisability from an independent direction.

One molecule is explained end to end, with additivity verified against the model's own output to 3.98e−13.

Every prediction ships a flag saying whether the model has seen chemistry like it. 22.2% of molecules are out of domain, and the flag predicts error:

Two details separate a useful flag from a constant: use unfolded Morgan substructure hashes (folded 2,048-bit fingerprints are degenerate at this corpus size — every bit is occupied), and cap the * points as methyl first (no corpus molecule carries an attachment point, so comparing raw makes every polymer novel by definition).

An honest negative result we chose to report: we tested the 5.97 M auxiliary corpus as a model input and it was worth +0.0012 — noise. An SVD embedding of it is a linear re-expression of the Morgan fingerprint the trees already see in full. We repurposed it as a trust signal rather than discarding it.

AISE2.0/
├── README.md                    this file
├── LICENSE                      MIT (code only — the data is not ours to license)
├── CLAUDE.md                    repo-wide orientation
└── Round 3/                     the finale — everything below is Round 3
    ├── CLAUDE.md                the operational contract: ground truth, rules, workflow
    ├── src/                     ★ source of truth for all modelling
    ├── submissions/             exported Kaggle notebooks + saved OOF/test vectors
    ├── experiments/LOG.md       ★ every scored run, dead ends included
    ├── context/                 the Kaggle pages (overview, rules, data, leaderboard)
    ├── documents/               reports, slides, the journey write-up
    ├── scripts/                 notebook linting, execution, diagram generation
    ├── .claude/                 skills, agents and hooks used during the sprint
    └── data/                    ⚠️ NOT IN GIT — see "Getting the data"

The competition data is deliberately not in this repository. Rule §4.b of the competition rules: "You agree not to transmit, duplicate, publish, redistribute or otherwise provide or make available the Competition Data to any party not participating in the Competition." (smile_r3.csv is also 314 MB, which exceeds GitHub's 100 MB per-file limit outright.)

Download from the competition data page into Round 3/data/:

Round 3/data/
├── train.csv               7,409 rows  (smiles, target, target_type)   long format
├── test.csv                4,940 rows  (id, smiles, target_type)
├── sample_submission.csv   a format illustration, not a template
├── PI1M.csv                995,800 polymer SMILES with * wildcards
├── smile_r3.csv            5,973,370 small-molecule SMILES (no wildcards)
└── base_line_model.ipynb   the host's official baseline

There is no system pandas, rdkit or lightgbm — use the project venv:

cd "Round 3"
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -r requirements-dev.txt

Every command below is run from Round 3/.

# Score a config — the gate every idea passes before it gets a submission slot
./.venv/bin/python -m src.cv --config lgbm            # ~377 s, the local reference
./.venv/bin/python -m src.cv --config lgbm_physics    # +0.0288, 4x the noise floor
./.venv/bin/python -m src.cv --config ridge_baseline  # host baseline equivalent, 4 s

# Full fit -> submission.csv  (~55 s for lgbm)
./.venv/bin/python -m src.predict --config lgbm --out submission.csv

# Validate before uploading — always
./.venv/bin/python -m src.check_submission submission.csv

# The judged deliverables
./.venv/bin/python -m src.invariance --config lgbm    # invariance certificate
./.venv/bin/python -m src.explain --config lgbm       # per-target TreeSHAP report

# The full five-model ensemble (long)
./.venv/bin/python -m src.run_ensemble --models lgbm,xgb,cb,mtnn,gnn --folds 10

# Regenerate the Kaggle notebook from src/
./.venv/bin/python -m src.export_notebook \
    --models lgbm,xgb,cb,mtnn,gnn --out submissions/final.ipynb

Runs are appended automatically to experiments/LOG.md. The harness prints a noise floor — a delta smaller than 2× it is not an improvement, and we said so out loud rather than claiming a win.

src/paths.py resolves the data at runtime (/kaggle/input/aisehack-2-0, then /kaggle/input, then local data/), so the exported notebook runs unmodified. Attach the Round-2 archive as a Kaggle Dataset if internet is off — otherwise the pipeline warns and runs archive-free rather than failing.

Round 3/submissions/ holds the full experimental record, not just the winner.

submissions/oof_and_preds/ holds saved OOF and test vectors for every base model, so a blending experiment costs seconds instead of a full rerun.

The full list is in journey.md. The five that generalise beyond this competition:

1. Fix your ruler before you build. Our first instinct was to improve a 0.903. The right move was to ask why 0.903 became 0.883. Every improvement measured with a broken validation is a coin flip.
2. Be suspicious of your biggest wins. The size of a measured gain correlated almost perfectly with the probability it was a leak: +0.069 was a leak, +0.042 was a leak, +0.008 was a leak. The real gains were +0.003 to +0.010 and arrived slowly.
3. Leaks have one shape — a value that is itself a model output, used as an input, closing a cycle back to the row's own label. Once we named the shape we found the rest quickly. And prove the leak exists before proving the guard removes it: a guard that passes without a demonstrable leak proves nothing.
4. Read where your data came from. The largest physics gain came from a 2020 DFPT paper and a 2016 dataset description; the second largest came from reading the host's own baseline notebook line by line. Neither was a modelling idea.
5. Only new information ever paid. Every purely derived feature failed. What worked: partner labels, a new model class, more data, truer structure, and an exact physical identity. If a feature is a deterministic function of columns the model already has, a good model already has it.

The physics and architecture choices are grounded in published work, principally from the Ramprasad group, whose datasets these targets originate from.

1. Chen, Kim, Batra et al. & Ramprasad, Frequency-dependent dielectric constant prediction of polymers using machine learning, npj Comput. Mater. 6, 61 (2020) — source of the ionic decomposition.
2. Huan, Mannodi-Kanakkithodi, Kim, Sharma, Pilania & Ramprasad, A polymer dataset for accelerated property prediction and design, Sci. Data 3, 160012 (2016) — the dataset of record; stores ε_electronic and ε_ionic as separate fields.
3. Gurnani, Kuenneth, Toland & Ramprasad, Polymer informatics at scale with multitask graph neural networks, Chem. Mater. 35, 1560 (2023) — polyGNN; source of the periodic graph.
4. Kuenneth, Rajan, Tran, Chen, Kim & Ramprasad, Polymer informatics with multi-task learning, Patterns 2, 100238 (2021) — the shared-trunk design, and the origin of this property set.
5. Lightstone, Chen, Kim, Batra & Ramprasad, Refractive index prediction models for polymers using machine learning, J. Appl. Phys. 127, 215105 (2020).
6. Gilmer, Schoenholz, Riley, Vinyals & Dahl, Neural message passing for quantum chemistry, ICML (2017).
7. Lundberg & Lee, A unified approach to interpreting model predictions, NeurIPS (2017); Lundberg et al., Nat. Mach. Intell. 2, 56 (2020) — SHAP / TreeSHAP.
8. Venkatram, Batra et al. & Ramprasad, Predicting crystallization tendency of polymers using multi-fidelity information fusion, J. Phys. Chem. B (2020) — ancestor of the partner features.
9. Liu et al., Open Polymer Challenge: Post-Competition Report, arXiv:2512.08896 (2025) — independently corroborates chain extension as a top-solution strategy.
10. Tran, Kim, Chen, Chandrasekaran, Batra et al., Machine-learning predictions of polymer properties with Polymer Genome, J. Appl. Phys. 128, 171104 (2020) — our external benchmark.