PEMFC Multiphysics Modelling

A proton-exchange membrane fuel cell is not governed by a single physical phenomenon. Its electrical response emerges from the interaction of electrochemistry, gas transport, water management, heat transfer, pressure dynamics, and operating controls.

For this reason, a useful engineering model must be multiphysics by design.

The objective is not to reproduce every microscopic mechanism. It is to retain the coupled phenomena that materially affect the quantities of interest while keeping the model identifiable, numerically robust, and computationally tractable.

This page describes the modelling structure I use for PEMFC digital-twin work. For a general introduction to the technology, see Proton-Exchange Membrane Fuel Cells.

The model is formulated at system and stack level rather than at pore or microstructure scale.

The principal modelled domains are:

• electrochemical conversion;
• membrane hydration;
• gas and vapour transport;
• liquid-water accumulation;
• anode and cathode pressure dynamics;
• thermal behaviour;
• purge-related operating states;
• electrical output.

The resulting model is intended for simulation, state estimation, engineering analysis, and eventually control-oriented applications.

The cell voltage can be represented schematically as

The reversible voltage depends on temperature and reactant partial pressures. The three principal losses correspond to:

• activation losses associated with reaction kinetics;
• ohmic losses dominated by proton transport through the membrane and electronic resistances;
• concentration or mass-transport losses at high current density.

The purpose of this decomposition is not only to reproduce a polarisation curve. Each term provides a physical connection between observable electrical behaviour and internal thermal, hydration, and transport states.

Membrane hydration is a central internal state because protonic conductivity depends strongly on water content.

A reduced model therefore tracks a membrane-hydration variable, typically written as a water-content state such as .

Its dynamics reflect several competing mechanisms:

• water production at the cathode;
• electro-osmotic drag from anode to cathode;
• back diffusion through the membrane;
• exchange with gas phases;
• removal through outlet flows.

Hydration directly affects membrane resistance and therefore the ohmic voltage loss.

A very dry membrane produces high resistance and poor performance. Excessive water, however, can contribute to flooding and transport limitations. The desired state is therefore not simply "more water".

Reactant pressures and compositions evolve dynamically with inlet flow, reaction consumption, outlet flow, humidification, and purge events.

A reduced control-volume formulation may be written generically as

The corresponding partial pressures follow from the gas inventory, volume, temperature, and equation of state.

These states are important because the electrochemical model depends on oxygen and hydrogen partial pressures rather than simply on externally imposed flow rates.

Water management requires distinguishing at least two forms:

• water vapour carried in the gas phase;
• condensed liquid water.

The transition between them depends on temperature, pressure, and saturation conditions.

A reduced liquid-water state can represent accumulation without attempting to resolve individual droplets or channels. Its role is to capture the effect of excess water on gas transport and operating behaviour.

This is particularly relevant for open-cathode or air-cooled systems, where thermal management, air supply, humidification, and water removal are tightly coupled.

Temperature affects nearly every subsystem:

• reaction kinetics;
• reversible voltage;
• membrane conductivity;
• saturation pressure;
• gas density;
• water condensation and evaporation;
• heat rejection.

A lumped thermal balance may be written as

Heat generation includes irreversible electrochemical losses and other internal dissipation. Heat rejection may include convection, forced airflow, conduction, and thermal exchange with surrounding structures.

The thermal model therefore closes several important feedback loops rather than acting as an independent post-processing calculation.

Dead-ended or periodically purged anode operation introduces a distinct operating timescale.

Between purge events, inert species and water can accumulate. A purge changes the gas inventory abruptly and therefore affects:

• hydrogen concentration;
• pressure;
• humidity;
• accumulated water;
• subsequent voltage behaviour.

A practical model can represent this through a continuous state describing accumulation since the last purge together with discrete purge events.

This naturally leads to a hybrid dynamical model containing both continuous states and event-driven transitions.

The important feature is not the existence of the individual submodels but their coupling.

For example:

1. current determines reactant consumption and water production;
2. water production changes membrane hydration and gas humidity;
3. hydration changes membrane conductivity;
4. conductivity changes the ohmic voltage loss;
5. losses generate heat;
6. temperature changes saturation pressure and reaction kinetics;
7. condensation changes liquid-water accumulation;
8. accumulated liquid water can restrict mass transport;
9. the resulting concentration loss changes terminal voltage.

A model that treats these effects independently can reproduce isolated curves while still failing dynamically.

A reduced state vector may contain quantities such as

depending on the model configuration.

Not every physical quantity should necessarily become an estimated state.

State selection is a compromise between:

• physical relevance;
• observability;
• parameter identifiability;
• computational cost;
• numerical conditioning;
• available measurements.

In the current state-estimation work, the estimated vector is deliberately smaller than the complete physical state.

Typical model inputs include:

• load current;
• inlet gas conditions;
• ambient conditions;
• cooling or fan command;
• purge command;
• operating set-points.

Typical measurable outputs include:

• stack or cell voltage;
• temperature;
• pressure;
• flow-related measurements where available.

The remaining internal quantities are model states rather than directly measured signals.

The implementation is developed in Python.

The main requirements are:

• explicit state definitions;
• deterministic model evaluation;
• clear physical units;
• stable numerical integration;
• bounded or physically constrained states;
• regression testing;
• reproducible parameter sets;
• separation between model equations and analysis code.

Scientific software quality matters directly here. A numerically fragile model can produce physically plausible plots while remaining unsuitable for state estimation, optimisation, or control.

A multiphysics model contains parameters of very different kinds:

• geometric parameters;
• thermophysical properties;
• electrochemical coefficients;
• transport coefficients;
• empirical correction terms;
• actuator and sensor parameters.

These should not be treated identically.

Some are known from geometry or material properties. Others require identification from data. Some are only effective parameters compensating for the reduced spatial resolution of the model.

The distinction must remain explicit because parameter fitting cannot turn an incorrect model structure into a physically valid one.

A full physical model is not automatically a better engineering model.

High spatial resolution may be useful for detailed design but can be counterproductive for real-time estimation or control.

Reduction therefore focuses on preserving the dominant input-output behaviour and physically meaningful internal states while removing unnecessary spatial detail.

The resulting model should be complex enough to represent the relevant couplings, but simple enough to analyse and execute repeatedly.

Experimental validation has not yet been completed for the current digital twin.

Before comparison with experimental data, the implementation can still be checked systematically through:

• dimensional consistency;
• conservation laws;
• limiting cases;
• monotonicity checks where physically expected;
• parameter sensitivity;
• numerical convergence;
• regression tests;
• comparison between equivalent formulations;
• internal consistency of coupled balances.

These checks are verification, not validation. They establish that the model is implemented consistently with its equations and assumptions; they do not prove that it reproduces the real fuel-cell system.

The digital twin is more than the forward multiphysics model.

A complete architecture also requires:

• measurement interfaces;
• parameter management;
• state estimation;
• uncertainty handling;
• comparison between model and system behaviour;
• persistent operating state;
• eventually diagnostic or control logic.

The multiphysics model forms the physical core of that architecture.

The next step is the estimation layer, where incomplete measurements are combined with model predictions to reconstruct internal states such as temperature, membrane hydration, and purge-related accumulation.

See also: State Estimation and Digital Twins.

The main principles I use are:

• retain physical meaning whenever possible;
• make assumptions explicit;
• separate measured quantities, states, and parameters;
• avoid fitting parameters that are not identifiable;
• test conservation and limiting behaviour;
• prefer reproducible parameter sets over manual tuning;
• distinguish verification from experimental validation;
• keep the model simple enough to understand and inspect.