Known limitations

What Thermal Architect does not model, what that does to an answer, and what to do about it. The first four are printed on every report. The theory manual states the assumptions behind everything else.

No three-dimensional airflow

What: Air is never solved as a flow field. A convection coefficient comes from a correlation for a named geometry (a plate, a channel, a heat sink, a jet) or is typed in; a fan-cooled enclosure's air is a fluid network of ducts and fittings with one mixed temperature per node.
Effect: Recirculation, bypass round a heat sink, hot spots behind a tall part and flow maldistribution between parallel channels are not predicted. The coefficient a correlation returns is an average over the surface.
What to do: Bracket the answer: run the study with the coefficient at the low and high ends of its plausible range. Where airflow detail decides the design, use CFD for the flow and bring the coefficients back here, or correlate the model to a thermocouple test.

Radiation is exchanged only where it is defined

What: Surfaces exchange radiation only through a radiation link or a radiation enclosure that you draw. Nothing is detected from geometry: there is no 3-D model to find which surfaces see each other. A single radiation link uses one emissivity and one view factor; the full grey-body exchange, with reflections between surfaces, is solved only inside a defined enclosure (validation case L01 shows the size of the difference).
Effect: Radiation you leave out is heat the model does not remove, which is conservative. Two grey surfaces joined by a plain link rather than an enclosure exchange too much, which is not.
What to do: For two or more surfaces that see each other, define an enclosure. Surfaces are grey and diffuse, with no participating medium and no specular reflection; the numerical rectangle view factor ignores third surfaces that would block the view.

No transient operating envelope yet

What: The operating envelope evaluates steady operating points only. Mission profiles (a sequence of conditions in time) can be described but are not yet solved. A single transient run of the model is available.
Effect: Worst cases that depend on timing - a soak followed by a power burst, a fan failure part way through a mission - are not found by the envelope, and a steady worst case may be pessimistic for a short pulse or optimistic for a long one.
What to do: Run the transient analysis on the specific sequence you are worried about, with the loads and boundaries set to that case.

Incompressible flow only

What: Fluid networks are solved as incompressible and steady: density is taken at each component's inlet state and held along it, and the flow has no inertia or storage. In a transient the loop is re-solved quasi-steadily as the walls warm. Compressors and compressible duct flow are refused with an error rather than solved.
Effect: Appropriate for liquids, and for gases where the velocity is low against the speed of sound and the pressure change is a small fraction of the absolute pressure. High-speed gas flow, choking, compression heating, pressure waves and water hammer are not represented. A gas component past about Mach 0.3, or with a pressure change past about 10 % of its inlet pressure, is warned.
What to do: Heed the compressibility warnings. Their thresholds (Mach 0.3, roughly 100 m/s in room-temperature air, and a 10 % pressure change) are engineering conventions, not physical limits. Split a long gas line into several components so density is re-evaluated along it.

Non-linear transients are semi-implicit, with no error control

What: Transients use backward Euler. For a network with radiation, contact, correlation-driven or temperature-dependent links the conductances are taken from the start of each step and not iterated within it. There is no adaptive step and no error estimate.
Effect: Stable at any step, but stability is not accuracy: backward Euler is first order, and a step that is large against the network's time constants smooths peaks and lags the response. The engine warns when the step is coarse for a load profile, or more than 0.3 of the time constant of a node holding a material share of the heat capacity; it has no error estimate beyond that.
What to do: Start near a tenth of the fastest relevant time constant, then halve the time step and confirm the answer stops moving.

Coolant loops are quasi-steady in transients and frozen in studies

What: A transient re-solves the coolant loop against the wall temperatures whenever a ported wall has moved by more than 0.01 K, but the coolant, its pipes and any reservoir store no heat, and flows follow the walls without inertia. Sweeps, optimisation, Monte Carlo, sensitivity and the advisor solve the loop once at the base design and hold it while the thermal parameters vary. The steady solve and the operating envelope re-solve the full coupling.
Effect: A loop whose own thermal mass matters (a large reservoir, long pipe runs) warms too quickly in a transient. A study whose parameters change the coolant flow or heat pick-up sees the base-design coolant.
What to do: Represent a reservoir or loop thermal mass with thermal nodes where it matters; check study results at their extremes with a full coupled solve.

Transient histories are thinned to 400 samples

What: Every step is computed, but each returned time history keeps at most 400 evenly spaced samples, plus the first and last. The maximum, minimum and mean of every series are computed from every step and returned separately, and study metrics use them.
Effect: A plotted curve can pass under a short peak that the reported maximum includes.
What to do: Read peaks from the reported statistics, not from the plotted history; shorten the run to see a peak's shape.

Heat pipe and vapour chamber start-up is not modelled

What: Heat pipes and vapour chambers are steady devices with capacity limits. In a transient their metal and charge store heat, but start-up, frozen-start and dry-out recovery are not simulated.
Effect: A transient through a heat pipe describes a pipe that is already running, and says nothing about getting it running. Beyond a limit the device is taken to carry its maximum plus envelope conduction, whereas a real pipe that dries out carries less: results past a limit are optimistic.
What to do: Check start-up against the manufacturer's data, and treat any operating point at or beyond a modelled limit as a failed design check rather than a temperature prediction.

Boards carry effective properties, not layers

What: A board is two plates, one per face, sharing the in-plane conductance equally and tied through the thickness. Which layer the copper is on, a single buried plane, a via field's detail and a pour's edges are not resolved.
Effect: The automatic mesh reads the junction about 1 to 2 % of its rise warm on typical boards (measured by uniform refinement), more where a part's theta_JB is near zero, which is warned. None of this is physical validation: the inputs (effective conductivity, theta_JB, the film coefficient) usually carry far more uncertainty than the discretisation.
What to do: Use the stack-up calculator for effective properties, re-solve with the board's refine factor at 2 to check a critical board, and bracket theta_JB and the film coefficient.

Imported layouts carry what the file states, not more

What: KiCad, Altium and IPC-2581 files hold no thermal conductivity, and often no dielectric thickness, plating thickness or component power. The import reads geometry and the stack-up; a dielectric whose material is not recognised is unresolved and blocks generation until assigned, and anything else is entered by the user or assumed explicitly and listed. Altium component bodies, plane splits and anti-pads are not read: parts are sized by their pads and internal planes are treated as solid copper. IPC-2581 negative feature sets on positive layers are not subtracted, and nets are not read: electrical connection is never taken as thermal.
Effect: Region conductivities are only as good as the stack-up and copper they came from; a solid-plane assumption overstates spreading near splits.
What to do: Review the stack-up table's assumptions, enter laminate and plating figures from the fabricator, and check large parts' sizes.

Contact conductance is an order-of-magnitude estimate

What: The Cooper-Mikic-Yovanovich and Mikic correlations disagree with each other by a factor of two and with measurement by more. The dominant unknown is the contact microhardness of the prepared surface.
Effect: With nominal bulk hardness (the default when no microhardness is given) the joint conducts better than a real one: an optimistic bound.
What to do: Supply a measured contact microhardness, or treat the joint's resistance as a range in a sensitivity study.

Built-in property data are nominal

What: Solid properties are handbook values near 20 to 25 °C, some with a temperature dependence. Outside a material's stated range a property is held at its value at the range limit and flagged. Fluid properties come from CoolProp. Interface materials, pumps and fans in the libraries are representative of their class, not of a product.
Effect: Alloy temper, ceramic grade, filler loading and a specific product's curve can each move a result by more than the solver's error.
What to do: Replace library values with supplier data before a result is relied on for a final design.

Laminar-turbulent transition is interpolated

What: Between Reynolds numbers of 2300 and 4000 (3000 for Gnielinski) friction factor and Nusselt number are blended linearly between the laminar and turbulent values.
Effect: Real transition depends on inlet conditions and disturbances; results in the band can be off by a factor of two, and the solver warns.
What to do: Move the operating point out of the band, or bracket it.