3D print duration regressor (AutoGluon) β€” 24-679 Fall 2026

Predicts how many minutes a 3D print will take from five job settings known before slicing.

Purpose

Built for Homework 2 (AutoML for classical machine learning) of 24-679 Designing with AI. The model estimates print duration so a designer can compare job settings without running a slicer. It is coursework, not a production planning tool.

Data origin and splits

Source: yennik16/3d-print-jobs-tabular, collected by a classmate for Homework 1. Thirty real print jobs were split 70/15/15 by job, then the training jobs were augmented.

Split Rows Distinct prints Contents
train 321 21 originals plus proportional_jitter and knn_interpolation variants
validation 4 4 original prints only
test 5 5 original prints only

The published splits are used unchanged. No print contributes rows to more than one split, and both holdouts contain original prints only; both properties are asserted in the notebook.

The effective training size is 21 prints, not 321 rows. Each training row is either one of 21 real prints or a synthetic variant of one, so the rows are not independent observations. Every uncertainty estimate below reflects that.

Features and target

Column Role Type Units
filament_g input float grams
infill_pct input int percent
layer_height_mm input float millimetres
nozzle_mm input float millimetres
use_category input categorical functional / novelty
print_time_min target float minutes

Excluded to prevent leakage: print_time_hms (the target restated as HH:MM:SS), long_print (a threshold on the target), and part_name (a per-job identifier). Provenance columns (source_id, parent_id, augmentation, is_augmented) are excluded as metadata.

Preprocessing

None beyond type coercion. Numeric columns are passed at their measured magnitudes and use_category is passed as a category; AutoGluon handles encoding internally. No scaling, imputation or outlier removal is applied, and no rows are dropped.

Training setup and search space

AutoGluon Tabular 1.6.1, problem_type="regression", eval_metric="mean_absolute_error", seed 24679. num_bag_folds=0 and num_stack_levels=0 are set so AutoGluon does not re-split the training rows internally, which would place synthetic siblings on both sides of a fold.

Search A β€” family portfolio. presets="medium_quality", 300 s budget, candidate families at their default settings, ranked on the 4 validation prints.

Search B β€” tuned random forest. 20 random trials, 300 s budget, over:

Hyperparameter Range
n_estimators 50 – 300
max_features sqrt, log2, 1.0
min_samples_leaf 1 – 8
max_depth None, 4, 8, 16

Ranges are narrow on purpose: 21 distinct prints cannot identify more.

Model selection

Candidates from both searches were scored the same way β€” mean absolute error on the four original validation prints β€” and the lowest was selected before the test prints were read.

Search Best candidate Validation MAE
A: family portfolio ExtraTreesMSE 239.9 min
B: tuned random forest RandomForest/T9 236.4 min

Selected: Search B β†’ RandomForest/T9 at 236.4 min.

The two searches finish within 3.5 minutes of each other on four validation prints, which is well inside the noise of a four-point measurement. The tuned configuration is selected on the rule fixed in advance, not because it is meaningfully better.

{
  "n_estimators": 247.0,
  "max_leaf_nodes": 15000.0,
  "n_jobs": -1.0,
  "bootstrap": 1.0,
  "max_features": "sqrt",
  "min_samples_leaf": 4.0,
  "max_depth": "None",
  "random_state": 0.0
}

Why the leaderboard was not used for selection

AutoGluon reports score_val for every model, but computes it differently depending on the search. With hyperparameter tuning active, each trial is scored on an internal split carved out of the training rows rather than on the tuning_data supplied. Those training rows contain synthetic variants, so a trial's internal holdout can contain a near-copy of a row it was fitted on, and the resulting scores are optimistic.

Re-scoring every Search B trial on the four original validation prints reorders them almost entirely. The configuration the leaderboard placed first (RandomForest/T2, leaderboard MAE 31.5 min) ranks 15 of 21 when measured on prints it never saw (262.9 min). Selecting on the leaderboard would have shipped a worse model while reporting a better number.

Anyone reusing this approach should score candidates themselves rather than trusting score_val when hyperparameter_tune_kwargs is set.

Feature importance

Permutation importance on the validation prints, measured as the rise in MAE when one column is shuffled (10 shuffle sets):

Feature Importance (min) p-value
layer_height_mm 145.2 0.000
infill_pct 74.3 0.001
filament_g 32.4 0.074
use_category 23.0 0.000
nozzle_mm -4.9 0.752

Layer height dominates, which matches the physics: print time is governed by how many layers the machine lays down, and that count scales inversely with layer height regardless of the mass being deposited. Filament mass β€” the input most people would reach for first β€” is third and not significant at the 5% level on this validation set.

Metrics

All errors are in minutes. Test set is 5 original prints.

Metric Model Mean-duration baseline
MAE 512.1 min 734.0 min
RMSE 970.3 min 1141.8 min
MAPE 87.1% 387.5%
RΒ² 0.120 -0.218

95% bootstrap confidence intervals (10,000 resamples of the 5 test prints):

Metric Estimate 95% CI
MAE 512.1 63.7 – 1343.9 min
RMSE 970.3 65.6 – 1673.2 min
MAPE 87.1 16.4 – 198.0%

These intervals are wide because five prints cannot pin down an error rate. Treat the point estimates as indicative, not precise.

Leave-one-print-out cross-validation over the 21 training prints, grouped by parent_id so no synthetic variant of a held-out print remains in the fitting data: MAE 195.5 min (std 165.7). This covers more prints than the test split and is the more stable of the two estimates.

Excluding the one extrapolation print, MAE on the remaining 4 test prints falls to 100.9 min, about a fifth of the full-test figure β€” that single print accounts for most of the headline error. MAPE stays high at 90% because the shortest test print runs 26 minutes, where a 79-minute miss is a large fraction of a small number. Reported as a diagnostic, not as the headline.

Known failure modes

  1. No extrapolation beyond the training range. Training durations span 27–1335 minutes. The test print ce_visor ran 2923 minutes and was predicted at
    1. Tree-based models average training targets, so no input can produce a prediction above the largest value seen in training. Predictions are meaningful only for prints expected to fall inside the training range.
  2. Unseen nozzle sizes. Training contains only 0.2 mm and 0.4 mm nozzles. A 0.6 mm nozzle appears in validation and is treated as an unknown category.
  3. Thin coverage of the input space. 21 prints cannot cover the combinations of mass, infill, layer height and nozzle that a real print farm sees.
  4. One printer, one operator, one filament workflow. Durations depend on machine, slicer profile and material; none of these vary in the data, so the model cannot account for them.
  5. Model selection is barely identifiable. The two searches' winners differ by 3.5 minutes on 4 validation prints, and the Search B trials span only 62 minutes end to end. At this sample size the ranking is largely noise, so the selected configuration should not be read as the best one available.

Limitations and ethical notes

The data records inanimate objects and machine settings; it contains no personal information. The practical risk is misplaced confidence: a designer trusting a duration estimate for a job unlike anything in the training set could schedule a print badly. The failure modes above are the boundary of reasonable use. Not suitable for costing, scheduling or any decision with a real consequence.

License

CC-BY-4.0, matching the source dataset's coursework terms. Provided for classroom use.

Hardware and compute budget

Environment macOS-15.5-arm64-arm-64bit-Mach-O
Python 3.13.5
AutoGluon 1.6.1
Accelerator CPU only; no GPU used
Search A 300 s
Search B 300 s (20 trials)
Cross-validation 60 s x 21 folds
Seed 24679

How to use

import huggingface_hub, zipfile, pathlib, pandas
import autogluon.tabular

path = huggingface_hub.hf_hub_download(
    repo_id="ssg1/2026-24679-print-time-autogluon-regressor", repo_type="model",
    filename="autogluon_predictor_dir.zip",
)
with zipfile.ZipFile(path) as archive:
    archive.extractall("predictor_dir")

predictor = autogluon.tabular.TabularPredictor.load("predictor_dir")

job = pandas.DataFrame([{
    "filament_g": 120.0,
    "infill_pct": 20,
    "layer_height_mm": 0.2,
    "nozzle_mm": 0.4,
    "use_category": "functional",
}])
print(predictor.predict(job, model="RandomForest/T9"))  # minutes

AI usage disclosure

Claude (Anthropic) was used to review candidate classmate datasets and identify leakage, draft the notebook structure following the course reference notebooks, and draft this card from values computed in the run. All modelling decisions β€” the choice of target, which columns to exclude, the grouped cross-validation scheme, and the reported metrics β€” were reviewed and verified against the notebook output. The data is a classmate's; no part of it was generated by AI for this model.

Author

Sohum Goel, Carnegie Mellon University, 24-679 Designing with AI, Fall 2026. Trained on a dataset by yennik16.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train ssg1/2026-24679-print-time-autogluon-regressor

Space using ssg1/2026-24679-print-time-autogluon-regressor 1

Evaluation results