Thermal-Crosstalk-Aware Scheduling and Model Predictive Dynamic Voltage and Frequency Scaling for Spaceborne Multi-Chip AI Clusters in Deep Vacuum: Lumped Multiphysics Modeling, Predictive Control, and Numerical Simulation

High-density heterogeneous artificial intelligence (AI) computing clusters onboard small satellites in Low Earth Orbit (LEO) face severe electrothermal constraints. Gas-mediated convective heat transfer is neglected in the adopted high-vacuum model (h ≈ 0), confining thermal rejection exclusively to internal solid conduction and non-linear Stefan-Boltzmann radiation to deep space (q ∝ T⁴). In multi-chip modules (MCM), heavy neural inference workloads induce severe lateral conductive coupling between adjacent silicon dies (thermal crosstalk). This paper investigates a model-based predictive thermal management strategy, denoted TCAS-DVFS (Thermal-Crosstalk-Aware Scheduling and Model Predictive Dynamic Voltage and Frequency Scaling), formulated from lumped thermodynamic principles for multi-die spaceborne accelerators. The proposed architecture integrates: (i) real-time state estimation via an Extended Kalman Filter (EKF) operating over a non-linear 5×5 Compact Thermal Model (CTM); (ii) a discrete linear time-varying Model Predictive Controller (LTV MPC-DVFS) that solves a condensed quadratic program via a warm-started ADMM-style splitting iteration, coordinating operating frequencies; (iii) a dynamic temperature-referenced radiative power-budget surrogate derived from orbital ephemerides; and (iv) a topological task allocator relaxed over the semidefinite cone S₊¹⁷ via Semidefinite Programming (SDP). In a three-orbit numerical co-simulation of a 500-km circular LEO scenario (17,004.0 s, 34,009 time samples, 4 processing dies), TCAS-DVFS produced zero observed junction temperature violations above the 85 °C threshold (0 s versus 200.5 s for binary Thermal Throttling at 85/75 °C and 14.0 s for Space-PID), attenuated peak inter-die thermal gradients by 77.6% (reducing from 9.42 to 2.11 K), and maintained an aggregate throughput proxy of 124.38 FPS with an energy-delay product proxy of 7.48 kJ·s/frame. Rather than asserting unconditional superiority, we characterize TCAS-DVFS as a multi-objective trade-off that enforces strict thermal containment and high throughput, whereas uncoordinated baselines like simple TCA achieve lower peak gradients (1.58 K) at the cost of lower throughput (116.16 FPS). Functional ablation indicates a dominant contribution from the predictive control formulation in damping gradients and bounding temperatures, whereas topological SDP scheduling provides limited observed benefits in compact arrangements (N = 4 dies) due to conductive symmetry. Sensitivity was evaluated under a 30-sample Latin Hypercube Sampling (LHS) stress-test envelope (±20% parameter variation across physical plant resistances and capacitances), yielding zero observed violations with an exact 95% Clopper-Pearson upper confidence bound of 11.57%, where die-to-radiator thermal resistance (R_r) exhibited the strongest monotonic association with T_max (Spearman ρ = +0.994). Indicative reliability proxy analyses demonstrate that electromigration yields no net life extension (Relative Mean Time to Failure [MTTF] proxy ratios of 0.96 and 0.81 on aggregate dies, with 1.52 and 1.33 on the hottest individual die), whereas cyclic thermal stress severity is substantially reduced (cycles above 2.0 K drop from 612 to 20). The reported evidence is limited to numerical simulation of the stated lumped-parameter model and does not constitute hardware or flight qualification.