Key Findings
A study published in Discover Electronics on 5 October 2026 demonstrates a physics-guided AI framework capable of predicting PEM fuel cell stack power with a coefficient of determination of 0.969. The model achieved a root-mean-square error (RMSE) of 0.295 watts and a mean absolute percentage error of 0.097 percent. When benchmarked against six other architectures, including LSTM, GRU, and Transformers, the chosen TS-Mixer architecture provided the best R-squared value at 0.963.
Technical Details
The researchers utilized a Time-Series Mixer (TS-Mixer) with parallel dense layers of 256, 128, and 64 neurons. Unlike typical physics-informed neural networks that use equations as penalty terms in the loss function, this framework uses first-principles quantities as engineered input features, including:
- Nernst open-circuit voltage
- Activation overpotential (Tafel approximation of the Butler-Volmer equation)
- Ohmic losses from effective membrane resistance
- Concentration losses from a limiting-current formulation
To account for degradation, the team added cumulative operating time and its product with current density as additional features. The model was trained using the Adam optimizer and evaluated against the IEEE PHM 2014 FC1 Ageing dataset, consisting of 65,535 sequential measurements.
Background & Context
The FC1 ageing test involved a 497-hour experiment where current density remained relatively stable (mean of 0.702 A/cm²), but measured voltage and power declined steadily. The researchers noted that a fixed-parameter electrochemical model alone correlates poorly with measured stack voltage (0.235). To avoid overstating predictive capability, the authors ensured the network was trained against independent experimental observables rather than values derived from the same equations used for the inputs.
Significance & Outlook
The ablation study showed that the physics-informed feature set improved the test R-squared by approximately 3.8 percentage points compared to predicting from elapsed time alone. While the feature-level approach is simpler to audit, the authors acknowledge it does not enforce physical consistency as a hard constraint, meaning predictions could deviate from electrochemistry outside the training distribution.
Source: https://bioengineer.org/physics-guided-ai-tracks-fuel-cell-ageing-with-watt-level-precision/
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
