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

Key points

The authors ran 729 material-extrusion 3D printing experiments across five materials, ABS, PETG, TPU, PLA, and recycled PLA, to build per-part greenhouse-gas emission data.
Each experiment is described by 22 material, process, and toolpath features, and the per-part emission value combines measured fabrication electricity with polymer-specific cradle-to-gate production factors.
A reinforcement-learning policy assigns weights among individual source predictors, source subsets, fully pooled sources, and target-only alternatives, and a validation gate deploys the transferred prediction only when it improves upon the supervised alternatives.
In leave-one-material-out evaluation, the method achieves a mean test RMSE of 5.2 g.CO2e per part, 20.0% lower than target-only training.
When only 5–10% of the target training data are available the RMSE is 32.5% lower than target-only fitting, and the gated method shows a negative-transfer rate of 17% versus 49% for pooling.
As limitations, the authors state that the validation-based deployment heuristic does not guarantee lower test error and that the five-material sample limits inference beyond the evaluated materials.

External research summaries. These are not HDATF publications or measured product results.