Aethermor solves steady or transient Fourier heat conduction on a Cartesian, cell-centred mesh and can couple one or more die faces to prescribed thermal boundaries. It is an architecture-stage reduced-order model. It does not solve fluid flow, radiation, transistor electrothermal feedback, package thermo-mechanics, or detailed interconnect/via geometry.
The transient equation is
[ \rho c_p \frac{\partial T}{\partial t} = \nabla!\cdot(\mathbf{k}\nabla T) + q''', ]
and steady state sets the time derivative to zero. Here T is temperature
[K], rho density [kg/m³], cp specific heat [J/(kg K)], k conductivity
[W/(m K)] or its diagonal tensor, and q''' volumetric heat generation
[W/m³]. The formulation follows the control-volume balance described by
Patankar, Numerical Heat Transfer and Fluid Flow.
For each cell P, steady energy balance is
[ \sum_f G_f (T_P - T_{N,f}) = Q_P, ]
where Q_P [W] is the integrated cell power. For an internal face between
cells P and N, normal conductivity is harmonically averaged:
[ G_f = \frac{k_h A_f}{d_{PN}}, \qquad k_h = \frac{2k_P k_N}{k_P+k_N}. ]
This produces the sparse system A T = b, solved with
scipy.sparse.linalg.spsolve.
The result reports the scaled infinity-norm residual, boundary heat flow, and
energy-balance error. A source-bearing system with no Dirichlet or Robin
temperature reference is singular and is rejected.
Each of the six faces can independently be:
- adiabatic/symmetry:
q'' = 0; - prescribed outward heat flux:
q'' = specified; - Dirichlet:
T_face = T_fixed; - Robin/convection:
q'' = h(T_face - T_infinity).
For a Robin face, conduction from the cell centre through half a cell and convection from the surface are resistances in series:
[ G_f = \frac{A_f}{d_n/(2k_n) + 1/h}. ]
This conductance is the only heat-removal path on that face. No ambient ghost
cell is used in convective mode. Edges and corners receive one contribution
for every exposed face. h=0 is exactly adiabatic.
Thermal input is a physical power map, not gates per numerical cell:
W/cell: integrated cell power;W/m³: multiplied by cell volume;W/m²: multiplied by cell plan area and deposited in a selected layer.
Therefore remeshing a fixed physical die and fixed power map preserves total power. Optional gate-energy models convert areal gate density [gates/m²], activity, frequency, and energy/switch into an areal heat input; they are not required by the thermal solver.
The compatibility transient path uses explicit conservative pairwise fluxes. Its time step is limited using the sum of directional diffusion rates. The steady architecture workflow does not estimate equilibrium by a fixed number of transient steps.
PackageStack represents the series path
junction -> die -> contact -> TIM -> contact -> spreader/IHS -> contact -> cold plate -> coolant.
Layer resistance is R=L/(kA) and area-normalized contact resistance is
R=R''/A. Spreading is a reduced-order correlation based on
Yovanovich, Muzychka, and Culham (1999), DOI 10.2514/2.6467.
When coupled to a meshed die, Aethermor uses external_resistance(), excluding
die conduction so it is not counted twice. The equivalent face coefficient is
h_external=1/(R_external A_die) and is applied only to the selected package
face.
R_ja, R_jc, theta metrics, and psi metrics are not interchangeable.
JEDEC theta_jc depends on a specified heat-flow/test configuration; Aethermor
does not claim its generic network reproduces that metric without matching the
standard. See the JEDEC JESD51-14 catalog description
and JESD51-12 guidance.
A design evaluation creates the configured model, solves the same steady system, and evaluates constraints. Maximum allowable power scales a fixed spatial power pattern and bisects until the active constraint is met. Results include the binding constraint and are re-solved at the returned point. Cross-material comparisons use one explicit device junction limit; material survival temperature is a separate additional constraint.
Local sensitivity uses central finite differences. Uncertainty propagation uses reproducible bounded Latin-hypercube samples and reports P5/P50/P95 plus a squared-correlation contribution ranking. Those correlations are screening indices, not a substitute for variance-based global sensitivity when inputs are dependent or the response is strongly non-monotonic.
aethermor validate currently demonstrates plane-wall accuracy, an
adiabatic/Robin generating slab, a multilayer resistance network, second-order
manufactured-solution convergence, and fixed-power mesh invariance. These are
numerical/analytical verification, not independent validation.
An executed pinned HotSpot v7.0 comparison supplies cross-code evidence for
synthetic Class-A cases constructed from shared canonical fields and Class-B
package trends. The finite source-observation mappings, post-output amendments,
and failures are disclosed. A matched Grade-A hardware holdout is explicitly
NOT RUN, so Aethermor has no hardware accuracy envelope and must not be used
for sign-off.
- orthogonal Cartesian cells;
- diagonal conductivity only; no off-diagonal tensor terms;
- piecewise-constant properties during one solve;
- no radiation or fluid-flow solution;
- no phase change;
- reduced-order package area/spreading correlations;
- one stable near-300 K bulk-silicon profile with property-level provenance; opt-in legacy materials remain provisional and cannot support release claims.