Learning When to Transfer: Reinforcement Learning for Carbon-Footprint Modeling in New Manufacturing Domains

Manufacturers often need to characterize a new material, process, or product before enough operating data are available to train reliable predictive models. Existing operations contain potentially useful knowledge for accelerating this characterization, but indiscriminate reuse can cause negative transfer when manufacturing domains differ. We propose a reinforcement-learning method for selective knowledge transfer. Given limited data from a new target domain, the learned policy chooses among predictors based on individual source operations, selected source subsets, pooled source and target data, or target data alone. A validation gate deploys the transferred prediction only when it improves upon a strong supervised alternative. The method is designed for learning manufacturing variables such as process energy and material consumption, which populate life-cycle inventories and support carbon accounting, as well as performance and quality outcomes. We demonstrate the approach through per-part greenhouse-gas prediction in material-extrusion additive manufacturing using 729 experiments across ABS, PETG, TPU, PLA, and recycled PLA. The response combines measured fabrication electricity with polymer-specific cradle-to-gate production factors, while 22 material, process, and toolpath features characterize each experiment. In leave-one-material-out evaluation, the proposed method achieves a mean test RMSE of 5.2 g CO$_2$e per part, 20.0% lower than target-only training and 10.3% lower than indiscriminate source-target pooling. The validation gate reduces the observed negative-transfer rate from 33% to 17%. These results demonstrate how manufacturers can selectively reuse data from existing operations to accelerate carbon-footprint modeling for a newly introduced material while limiting harmful transfer.