ML-guided polymer discovery loop

Context

polymer-ml-lab is a public, notebook-led study. It is not an industrial R&D deployment and it does not use proprietary lab measurements. The premise is scientific and didactic: use ML to shrink a theoretical search space before anyone spends experimental budget.

This is the clearest public bridge from physics engineering to data systems: constraints first, then measurement (even if simulated), then a decision.

Problem

Formulation spaces are large. Tensile strength, elongation, and thermal resistance pull in different directions. A team that tests at random burns time. A team that only follows folklore may never see the Pareto surface. The project asks how to propose a small set of candidates that are physically plausible and worth confirming.

Architecture

The documented loop:

1. Define a constrained formulation sandbox (base, additives, process variables, mass balance ≈ 100 wt%).
2. Generate pseudo-lab data with physics-inspired equations plus controlled noise.
3. Train surrogates (linear baselines vs Random Forest), including a low-data regime.
4. Generate candidates with constrained search and Upper Confidence Bound (UCB) scoring.
5. Simulate measuring those candidates, append, retrain, watch the Pareto front.

Code is split between ordered notebooks (01 data generation, 02 modeling, 03 optimization) and src/data_generation.py so sampling is not trapped in a notebook.

Data / inputs

There is no confidential polymer dataset. Bounds, samplers, and a lab-simulation helper live in code. Generated tables can be cached locally and are gitignored by default. Trends the README expects: filler raises tensile strength and hurts elongation; plasticizer does the opposite; temperature has a bell-shaped optimum.

Engineering decisions

• Physics in the generator, not as a vibe. Mass balance and process limits are encoded so the model cannot recommend nonsense formulations as freely.
• Baselines before clever models. Linear models sit next to Random Forest so “ML helped” is comparable, not theatrical.
• Uncertainty-aware reading of low-data regimes. The write-up emphasizes not trusting overconfident recommendations when samples are few.
• UCB as the decision policy. Exploration versus exploitation is explicit: the loop is allowed to visit novel regions, not only the current best.
• Notebooks as lab books. Each notebook ends with notes pointing to the next step. Reproducibility is the product.

Implementation

Environment: NumPy, pandas, scikit-learn, matplotlib, seaborn. Run notebooks in order, or call python -m src.data_generation. A helper strips notebook outputs before version control so diffs stay readable.

Challenges

• Synthetic data only approximates a lab. The README is clear: real measurements should replace or fine-tune the same notebooks (Gaussian processes are listed as a future surrogate).
• Optimization is heuristic search, not a full Bayesian optimization stack. That is an honest limitation, not a hidden one.
• Extra industrial constraints (cost, aging, rheology) are named as missing, not silently assumed.

Results

From the project README, as a synthetic study, not a factory trial:

• Random Forest surrogates stayed robust when data was throttled to about 20–30 samples.
• The UCB loop narrowed the Pareto front in roughly three iterations.
• On the order of the top ~5% of theoretical candidates would be sent to the (simulated) lab.

I will not translate those figures into “we cut lab cost by X% at company Y.”

What I learned

Measure before optimizing: even in a toy loop, you compare a linear baseline and you watch uncertainty. Systems > tools: the value is the closed loop (generate → score → choose → measure), not the choice of forest versus network.

Grain matters here too. A formulation vector is not a recommended experiment until it has survived constraints and a policy. Mixing those grains is how a notebook becomes an overconfident slide.

Stack

Python, NumPy, pandas, scikit-learn, matplotlib, seaborn, Jupyter.

Repository

github.com/carlosdcorona/polymer-ml-lab