Robust feasible-domain modeling for resource-constrained spaceborne heterogeneous high-performance computing

A spaceborne heterogeneous computer must finish sensing, preprocessing, artificial-intelligence inference, compression, and data return within one mission window. Processor throughput alone is an incomplete guide because task placement simultaneously determines computation, transfer, energy, peak power, thermal accumulation, and reliability exposure. We developed a reproducible feasible-domain workflow that screens compatibility, transient memory, deadlines, period, peak power, repeated-cycle temperature, and segment reliability before multi-objective selection. Its executable model combines a directed task graph, task-dependent Roofline capability, mapping-induced communication, baseline-increment energy, a first-order RC recurrence, and bounded perturbations. Direct enumeration of a synthetic six-task case reduced 4096 CPU/GPGPU/NPU/FPGA mappings to 18 after timing and capacity checks and to 15 after thermal screening. Compared with the nominal energy-oriented mapping, robust selection used 1.51% more energy and shortened makespan by 122.701 ms. No deadline violation was observed in 1000 seeded replay samples at the declared perturbation strength; the Wilson 95% upper bound was 0.383%. Because the inputs are theoretical model parameters without hardware measurements, the study supports reproducible design screening and makes no claim about flight-system performance.