A new strategy called CA‑FCS chooses which phones train a shared model by looking at how much carbon each device will use, not just how much battery it will drain. The method builds an energy model from real tests on Google Pixel phones and adds hourly forecasts of the grid’s carbon intensity. By estimating each client’s carbon credit, CA‑FCS picks the users that will emit the least.

To keep the global model strong, CA‑FCS limits how many times any phone can train – at least once and at most twenty times – and it never asks the same phone in two consecutive rounds. This keeps data diversity and avoids bias toward only low‑emission users.

Tests on five real datasets, with data divided evenly or in heavily skewed ways, show that CA‑FCS lowers client‑side carbon by up to 46% compared with random selection. Accuracy remains similar across all settings, though in some highly uneven data cases the score can dip by up to 3% because some low‑emission clients are temporarily skipped.

The experiments were run in realistic mobile federated learning environments where phones only participate when they are charging and connected to Wi‑Fi, just like a normal smartphone user would experience.

CA‑FCS therefore offers a practical way to train distributed models on mobile devices while keeping the environmental cost down, although the participation limits may restrict extreme use cases such as aggressive over‑training.