Often not — and in our own tests, not once. Across four independent evaluations we ran on public datasets, the physics term did not improve accuracy in any of them. Our networks have beaten gradient boosting. When we removed the physics, they still did.
In our tests the network wins — but not because of the physics. We ran a physics-informed network and a gradient-boosted baseline to a proper training budget on synthetic silicon-carbide power-module data. The network beat XGBoost on all four degradation measures, on both test conditions, in every one of ten runs. We then removed the physics term, and the network still won. The physics made it reach that result faster, not more accurately. This was the one evaluation we used to score as a win for physics, and it is not one. The literature in this area reports wins almost exclusively, which makes it a poor guide to what will work on your asset. What follows is our scorecard, the datasets behind it, and where we still think physics earns its place.
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 lostIGBT reproduction (NASA PCoE accelerated aging). Best performer: a standard LSTM, across four folds. The physics-informed variant was worse.
Physics lostNASA 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 lossSimulated SiC devices. The neural model beat the gradient-boosted baseline on all four measures, in every one of ten runs. Remove the physics term and it still wins: at a correct training budget the physics advantage was 0.6% and not statistically significant. The architecture, not the physics.
Physics did not winNone of the 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.
On simulated SiC devices we re-ran the comparison at a proper training budget. The physics stopped mattering.
Our earlier read was that physics was load-bearing here: against a baseline with the physics removed, the model roughly halved multi-target error. Re-run at the correct training budget, that advantage fell to 0.6% and was not statistically significant. The physics term made the model reach its result faster, not more accurately.
The win that survives belongs to the architecture. The neural model beat a tuned gradient-boosted baseline on all four degradation measures, on both test conditions, in every one of ten runs. That is worth something — but it is not evidence for physics-informed learning, and we will not present it as such.
Two further limits stand either way. The data is physics-based synthetic, and the generator encodes the very relationship being tested — so the result speaks to method behaviour, not to measured devices. And the held-out band was used both to select the model and to run the comparison, so the 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.
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.
A read-only assessment runs on data you already have, before any hardware conversation. You see the record; you decide what it is worth.
