Manufacturers often need to characterize a new material, process, or product before enough operating data are available to train reliable predictive models for this new, target domain. Existing manufacturing operations contain potentially useful knowledge to accelerate this characterization, but indiscriminate reuse of data from existing source domains can cause negative transfer in learning of the new, target domain when the manufacturing domains differ sufficiently. This work proposes a reinforcement-learning method for selective knowledge transfer, where its reward scores each opportunity to transfer against supervised alternatives, so negative transfer is avoided and controlled during forward prediction instead of being diagnosed afterwards. Given limited data from a new target domain, the learned policy chooses among predictors based on individual source operations, subsets of the sources, or the supervised alternatives: fully pooled sources, or target data alone. A validation gate deploys the transferred prediction only when it improves upon the supervised alternatives. The method is designed for learning variables supporting carbon accounting, as well as performance and quality outcomes. We demonstrate the approach by quantifying the embodied greenhouse-gas emissions of parts produced via material-extrusion additive manufacturing, using data from 729 experiments spanning five materials: 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, this method achieves a mean test RMSE of 5.2 g CO2e per part, 20.00% lower than target-only training (32.5% lower when only 5-10% of the target training data are available) and 10.30% lower than indiscriminate source-target pooling.
Figures & tables
Figure 1: Manufacturing-development context and transfer-policy workflow. Top: conceptual links between process development, supporting data, and manufacturing domains. Bottom: target-domain evidence informs an actor that assigns weights to candidate predictors. A critic supports reward-based policy training, and a validation gate selects between the policy mixture and the stronger supervised alternative.
Figure 2: Experimental setup and characterization of the five material domains:(a) material-extrusion printing; (b) estimated density curves of the per-part carbon response defined in Eq. 1 , in kgCO2e/part ; and (c) material descriptors projected onto the first two principal components, PC1 and PC2.
Test RMSE ( kgCO2e/part )
Negative-transfer rate (%)
b
AC + gate
BC + gate
Raw AC
Pooled
Target only
AC + gate
BC + gate
Raw AC
Pooled
0.05
0.0072
0.0076
0.0076
0.0075
0.0113
4
11
11
15
0.10
0.0065
0.0067
0.0069
0.0069
0.0090
2
16
22
20
0.20
0.0058
0.0059
0.0061
0.0062
0.0069
18
29
33
38
0.40
0.0047
0.0049
0.0051
0.0054
0.0053
22
33
38
51
0.60
0.0044
0.0044
0.0047
0.0053
0.0046
27
31
42
69
Table 1: Held-out test RMSE ( kgCO2e/part ) and negative-transfer rate by target fitting budget b . Negative-transfer rate is 100 times the number of cases in which a method’s test RMSE exceeds target-only test RMSE, divided by the number of evaluated cases. Each budget contains 5 materials ×11 repetitions =55 cases against each budget; overall results contain 5×7×11=385 cases. RMSE entries are means; the Avg. RMSE and Test R2 entries report mean ± standard deviation across all cases. AC: actor–critic; BC: behaviour cloning; Pooled: all sources + target.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Budget b
0.05
0.10
0.20
0.40
0.60
0.80
1.00
Target fitting observations
4
7–8
14–15
29–30
43–45
58–60
72–75
Fixed validation observations
approximately 29–30 at every budget
Appendix
Table 2: Approximate target fitting-set sizes. Ranges reflect the difference between the 145-sample and 149-sample material domains and integer rounding.
Method
Transfer basis
Rule
Purpose
Uniform average
Prediction
Equal-weight average of all source experts
Tests whether naive source ensembling is sufficient.
Similarity-weighted
Prediction
Inverse weighting by dmat(k)
Tests whether material proximity alone identifies useful sources.
Best single source
Prediction
Source with lowest error on At(b)
Hard source selection using limited target fitting evidence.
Adaptation simplex
Prediction
Nonnegative convex stacking of source predictions fitted on At(b)
Best tree model fitted on all source training data plus At(b)
Indiscriminate supervised transfer reference.
Appendix
Table 3: Evaluated baselines and the aspect of transfer tested by each method.
Reward Component
Formulation
Purpose
Negative validation
−NRMSEpol
Penalizes prediction error
Improvement bonus
+1.5[NRMSEref−NRMSEpol]+
Rewards improvement over reference
Underperformance penalty
−3.0[NRMSEpol−NRMSEref]+
Penalizes worse performance than reference
Expert-gap penalty
−0.75[NRMSEpol−NRMSEbest]+
Penalizes worse performance than the validation-optimal mixture
Appendix
Table 4: Reward components and their purposes.
Method
Val. R2
Val. RMSE
Test R2
Test RMSE
Actor–critic + gate
0.8580±0.1496
0.0044±0.0015
0.7757±0.2151
0.0052±0.0021
Behaviour cloning + gate
0.8303±0.2777
0.0047±0.0020
0.7264±0.2130
0.0053±0.0022
Raw actor–critic
0.8064±0.2535
0.0053±0.0020
0.7195±0.1893
0.0056±0.0022
All sources + target
0.8098±0.2679
0.0050±0.0020
0.6633±0.2654
0.0058±0.0023
Target only
0.6703±0.3684
0.0066±0.0033
0.4929±0.6609
0.0065±0.0033
Adaptation simplex
0.4851±0.6410
0.0077±0.0036
0.5005±0.3311
0.0077±0.0039
Appendix
Table 5: Overall performance across held-out target materials, fitting-data budgets, and subset draws. RMSE is in kgCO2e/part ; values are mean ± standard deviation.
Figure 3: Test RMSE versus target fitting budget for (a) PETG, (b) ABS, (c) PLA, (d) rPLA, and (e) TPU. The budget b is the fraction of the target training partition used for fitting; the separate validation partition remains fixed. Curves compare the complete policy, its ablations, target-only fitting, and all-sources-plus-target pooling. RMSE is expressed in kgCO2e/part . Lines show mean test RMSE; shaded bands show standard deviation.
Figure 4: Deployment evidence: (a) negative-transfer rate relative to target-only fitting; and (b) frequency with which the validation gate replaces the actor–critic prediction with the stronger supervised model.
Figure 5: Source-expert test RMSE against standardized material distance for representative source–target pairs; statistics over all 20 directed pairs are given in Appendix E .
Figure 6: Failure analysis using (a) absolute prediction error and (b) mean absolute error by source-expert disagreement group.
Method
RMSE ( kgCO2e/part )
Neg. transfer (%)
p (RMSE)
p (neg. transfer)
Actor–critic + gate (proposed)
0.0052 [0.0051, 0.0053]
17 [14, 20]
–
–
Behaviour cloning + gate
0.0053 [0.0053, 0.0054]
28 [25, 30]
0.13
4×10−5
Raw actor–critic
0.0056 [0.0055, 0.0057]
33 [30, 36]
1×10−9
2×10−11
All sources + target
0.0058 [0.0057, 0.0059]
49 [45, 52]
4×10−30
5×10−35
Target only
0.0065 [0.0063, 0.0067]
–
3×10−31
–
Appendix
Table 6: Held-out results with 95% intervals (385 cases; intervals resample cases within material–budget cells) and paired tests against the proposed method (Wilcoxon signed-rank for RMSE, McNemar for negative transfer). Intervals and p -values are case-level and treat repetitions as replicates for the evaluated materials.