Do physics-informed neural networks outperform XGBoost for remaining useful life prediction?

Frequently not. Across four independent evaluations Choir ran on public datasets, the physics-informed model won once. On real commercial cells from the MIT/Stanford Severson-Attia dataset a plain data-driven model placed first and the full physics-informed model placed fourth. On the NASA IGBT accelerated aging dataset a standard LSTM beat it across four folds. On the NASA battery ablation the physics loss terms moved RMSE by 0.0002 Ah, which is negligible. Only on simulated SiC devices did the physics-informed model win.

Why does XGBoost do so well on remaining useful life problems?

Because prognostics features are usually tabular, and tree ensembles remain strong on tabular data. Grinsztajn, Oyallon and Varoquaux demonstrated this at NeurIPS 2022. A physics-informed network losing to gradient boosting on a tabular RUL problem is the expected prior rather than a surprising result.

Is the published literature on physics-informed machine learning reliable?

It should be read with care. McGreivy and Hakim, publishing in Nature Machine Intelligence in 2024, surveyed 82 articles applying machine learning to fluid-related partial differential equations and found 76 claimed to outperform against 4 claiming to underperform, with 79% of the outperforming claims measured against a weak baseline. They concluded the literature is not currently a reliable source for evaluating whether these methods succeed.

When is a physics-informed model actually worth using?

The defensible case is extrapolation — predicting outside the operating range an asset has already run, where a purely statistical model has no support. Within the observed range, a well-tuned statistical model is usually competitive and easier to maintain. Choir treats this as a hypothesis under test rather than a settled claim, and reports evidence in both directions.

Why would a company publish evaluations where its own models lost?

Because an unaudited claim in this field is worth very little. The published literature reports wins almost exclusively, so a vendor scorecard showing only wins carries no information. Choir publishes the losses so that the one win has meaning, and so that any figure quoted from the site can be checked against a dated report of record.

Choir^™ CorpBook an assessment

Home — Answers — PINN vs XGBoost

Technical question · Prognostics

Do physics-informed neural networks beat XGBoost for remaining useful life?

Often not. Across four independent evaluations we ran on public datasets, the physics-informed model won once. On real commercial cells a plain data-driven model placed first; on IGBTs a standard LSTM beat it; and on the NASA battery ablation the physics loss terms moved error by a negligible amount.

The short answer

We are a physics-informed company publishing a mostly negative result about physics-informed models, so it is worth saying why. The literature in this area reports wins almost exclusively, which makes it a poor guide to what will work on your asset. Our position is that the technique should be chosen per asset class on evidence, and that the evidence is only worth anything if the losses are published alongside the wins. What follows is our scorecard, the datasets behind it, and the limits on the one evaluation we won.

The scorecard

Four evaluations. Physics won one.

Each row names the dataset. Each traces to a dated internal report; the figures behind them are available under diligence.

Real commercial cells (MIT/Stanford, Severson–Attia LFP). Best performer: a plain data-driven model. The full physics-informed model placed fourth. Reported as found.

Physics lost

IGBT reproduction (NASA PCoE accelerated aging). Best performer: a standard LSTM, across four folds. The physics-informed variant was worse.

Physics lost

NASA battery ablation. The physics loss terms did essentially nothing — they moved RMSE by 0.0002 Ah. Whatever improvement the study showed came from elsewhere in the model.

Not the physics loss

Simulated SiC devices. Physics-informed won, and the result held across ten seeds. On synthetic data — see the limits below, which are substantial.

Physics won

One evaluation in four. We keep the scorecard on the Evidence page in the same form we show customers, and it has not been edited to look better than it is.

The one we won

And why we are not leaning on it.

On simulated SiC devices, physics was load-bearing — and the effect was seed-stable.

Against a baseline with the physics removed, the physics-informed model roughly halved multi-target error and won on every one of ten seeds. Against a tuned gradient-boosted baseline on a bracketed held-out temperature band, it held on two of four targets; the other two were practical ties.

Two things stop this being the headline. The data is physics-based synthetic, and the generator encodes the very relationship being tested — so the result demonstrates that the method recovers physics that was injected, not that it works on measured devices. And the held-out band was used both to select the model and to run the comparison, so the comparison margin carries an optimistic bias.

The margins stay under diligence until real-device validation exists. We do not say “validated on measured SiC,” and we do not say “validated on your modules.”

Why gradient boosting is a hard baseline, not a weak one.

Grinsztajn, Oyallon and Varoquaux showed at NeurIPS 2022 that tree ensembles still outperform deep learning on typical tabular data. Prognostics features are usually tabular. So a PINN losing to XGBoost is not an embarrassment — it is the prior, and any paper reporting otherwise should be read carefully.

The reporting problem is measurable. McGreivy and Hakim, in Nature Machine Intelligence in 2024, surveyed 82 articles applying machine learning to fluid-related PDEs: 76 claimed to outperform and 4 claimed to underperform, and 79% of those claiming to outperform compared against a weak baseline. Their conclusion was that the literature is not currently a reliable guide to whether the methods work.

Krishnapriyan and colleagues showed at NeurIPS 2021 that physics-informed networks fail on problems whose plain neural counterparts are straightforward. None of this makes the approach useless. It makes unaudited claims about it worthless.

So when does physics earn its place?

Our current answer, stated as a hypothesis.

Not for interpolation. Possibly for extrapolation — and that is the test we are still running.

Where an asset has run across the conditions you care about, a well-tuned statistical model is hard to beat and usually easier to maintain. The case for encoding governing equations is about predicting outside the range the asset has run before, which is exactly where a purely statistical model has nothing to stand on.

Our own evidence there is mixed and we report it that way. On a temperature-extrapolation study using published Sandia cell data, a variant with competing high and low temperature degradation branches beat a gradient-boosted baseline on a held-out cold condition and roughly tied on a held-out hot one. That is the most encouraging real-cell evidence we have — and it is not yet seed-verified or pre-registered, so we hold it as suggestive rather than proven.

Meanwhile a strict extrapolation test elsewhere in our SiC work went the other way: a gradient-boosted baseline beat the physics network on both extrapolation experiments. We report that too. The reason we can be trusted on the wins is that we publish the losses.

What this does not show

Four evaluations is a small sample, and three of the four use public datasets with known limitations.

The SiC win is on synthetic data whose generator encodes the tested relation. It is not device validation.

The Sandia extrapolation result is not seed-verified or pre-registered. Suggestive only.

Choir has no deployed customer reference. Everything here runs on public or simulated data.

See the same discipline on your own assets.

A read-only assessment runs on data you already have, before any hardware conversation. You see the record; you decide what it is worth.

Book an assessment See the evidence

Early access · software-first · every number traces to a dated report