EVOLVE: learned volume compression with variable-rate encoding across a cross-domain scientific database

arXiv·low signal

EVOLVE (arXiv 2607.18187) targets large-scale scientific simulations that generate volumetric data faster than storage and network bandwidth improve, using learned compression with variable-rate encoding trained across a cross-domain database rather than per-dataset. Variable-rate encoding lets one model serve many fidelity budgets without retraining. Infrastructure-adjacent to agent work, but the 'one learned codec, many rate points' pattern is directly reusable for agent memory and artifact stores.