Can machine learning sort kiln-dried hemlock by final moisture? What schedule and conditioning add

Introduction

Every kiln operator knows the uncomfortable truth behind a "dry" charge: the boards that come out do not all share the same moisture content. Some are over-dried, wasting energy and risking degrade; others are still wet and must be re-dried, sorted out, or sold at a lower grade. Reducing this spread is one of the oldest goals of kiln drying, and most of the tools used to pursue it are drying schedules (the planned sequence of temperature and humidity steps) and conditioning (a final humid phase intended to even out moisture and relieve drying stresses).

A research group at the University of British Columbia, together with FPInnovations, asked a practical question: how much of the final moisture of an individual board can be explained by what the board looks like going into the kiln (its initial moisture, weight and density) and how much by what the operator does to it (schedule, conditioning, post-drying storage)? They used a machine learning method, gradient boosting, both to predict final moisture as a number and to sort boards into "acceptable", "over-dried" and "under-dried" classes. This article summarises that study, which was published in 2023 in the open-access journal Polymers, and then separates what the authors found from what a sawmill might reasonably take from it.

Research methodology

Material and preparation

The material was second-growth western hemlock (Tsuga heterophylla), part of the hem-fir commercial group, sourced from a sawmill on Vancouver Island, British Columbia. The researchers started with 96 green "baby square" timbers measuring 116 x 116 mm and 3.96 m long, grade II (standard) or better. After removing 100 mm from each end, every timber was cut into four kiln specimens of 900 mm and five 25 mm thick "cookies". That gave 384 kiln specimens and 480 cookies. Six kiln specimens were discarded, leaving 378 specimens, which were divided into nine drying batches of 42 pieces each. Batches were assembled so that the spread of initial moisture content and basic density within each batch was as small as possible, which helps make the batches comparable.

The cookies were used to measure initial moisture content and basic density with standard laboratory methods (the authors cite Kollmann and Skaar). Each kiln specimen was weighed before drying, and its end faces were sealed with polyvinyl acetate to limit end drying. Final moisture content was calculated afterwards from reweighing, using oven-dry-equivalent weight estimates.

Kiln and schedules

Drying took place in a conventional heat-and-vent laboratory kiln of 0.73 cubic metres at FPInnovations in Vancouver. The same aluminium stickers (19 mm thick) were used in every run, with the specimens arranged in six rows and seven columns. The target final moisture content was 12 percent, chosen to match average equilibrium moisture conditions in Japan, the intended market. Three schedules were used:

• Control schedule: based on industrial British Columbia kiln practice, with eight time-based steps totalling 180 hours, followed by a moisture-based step to the 12 percent target.
• Schedule I: six time-based steps (about 144 hours) plus a moisture-based step, ending at a dry-bulb temperature of 78.8 C, with a faster fall in the equilibrium moisture content setting than the control.
• Schedule II: five time-based steps plus a moisture-based step, ending at 85.0 C. This was the most aggressive schedule. Dry-bulb temperature was kept below 93 C to avoid honeycomb.

On top of the schedule, two optional post-treatments were varied. Conditioning was a 12-hour step at 71.7 C dry bulb and 66.7 C wet bulb (equilibrium moisture content 12.3 percent). Post-storage was 168 hours at 20 C and 65 percent relative humidity. Nine runs combined these factors (for example, Schedule I with conditioning and no storage, Schedule II with neither, and so on). After each run the timber cooled for 12 hours inside the closed kiln.

Machine learning set-up

The modelling used TreeNet gradient boosting with CART trees as base learners, for both regression (predicting final moisture content as a number) and classification. There were six inputs: initial moisture content, initial weight, basic density, drying schedule, conditioning (yes/no) and post-storage (yes/no). To test whether the process variables add anything, the regression was also run with only the three wood attributes.

For classification, final moisture content below 10 percent counted as over-dried, 10 to under 19 percent as acceptable, and 19 percent or more as under-dried. Key settings included a maximum of 2000 trees (the optimum found was 550 trees for the six-input regression and 214 for the three-input one), at most 12 terminal nodes per tree, a learning rate of 0.01 and a subsample fraction of 0.3. Hierarchical variable clustering and correlation analysis were used to look at relationships among inputs and final moisture. Performance was reported for training and test partitions with R-squared, RMSE, MAD and MAPE for regression, and per-class accuracy for classification.

Key findings

Schedule effects on the final moisture

The schedules gave quite different outcomes. Schedule I produced the highest mean final moisture (15.59 percent) and the greatest variation (5.16 percent as reported). The control schedule, with its longer drying time, gave the lowest mean (10.87 percent) and the lowest variation (2.19 percent). Conditioning was associated with a lower standard deviation of final moisture (4.31 percent with conditioning versus 5.47 percent without), and post-storage also lowered it (4.54 percent versus 5.07 percent).

Which initial properties matter

Clustering showed three groups of variables, with initial moisture and initial weight the most similar pair and basic density the least similar to final moisture. Correlations between initial properties and final moisture were modest and varied from run to run. Initial weight was the best single correlate in eight of nine runs, ranging from about 0.38 to 0.75; initial moisture correlated between 0.15 and 0.62; and basic density between 0.12 and 0.60. The authors note that these correlations were weaker than in their earlier single-schedule study without post-drying treatments, which is what one would expect when more process variation is mixed in.

Regression: not good enough

With all six inputs, the regression model achieved an R-squared of 73.86 percent on the training data but only 44.81 percent on the test data (test RMSE 3.61). With only the three wood attributes it fell to 56.80 percent training and 32.10 percent test (test RMSE 4.00). So adding the schedule, conditioning and storage variables clearly helped, but the gap between training and test performance signals overfitting, and the authors judged the regression unsatisfactory for predicting an individual board's final moisture.

In the six-input model, relative importance was led by initial weight (set to 100), then initial moisture (92.6), basic density (84.4), schedule (63.3), post-storage (45.0) and conditioning (41.0).

Classification: more useful

Classification worked better. Overall accuracy was 76.19 percent on training data and 69.05 percent on test data. For the test data, the correct-classification rates were 58.62 percent for over-dried boards, 71.64 percent for acceptable boards and 67.31 percent for under-dried boards. Put differently, roughly 31 percent of the 378 specimens were misclassified in the test results, with the over-dried class the hardest to catch. Relative importance again put initial weight first (100), followed by initial moisture (90.0), basic density (71.8), schedule (53.7), post-storage (43.7) and conditioning (32.7).

The authors concluded that a classifier of this kind performs acceptably and could support sorting before and after drying, for instance in deciding where to set breakpoints in a dry-sort-re-dry strategy.

Practical applications for sawmills

Points that come from the paper

• Initial weight of a board was the most informative of the three wood attributes for final moisture, and weighing is quick and non-destructive. The authors highlight this as relevant for sawmill implementation.
• Process variables (schedule, conditioning, post-storage) improved the models' accuracy compared with wood attributes alone, so a model that ignores how the charge was dried will be less reliable.
• Conditioning and post-drying storage were both associated with a narrower spread of final moisture in this dataset.
• Sorting boards into over-dried, acceptable and under-dried classes was feasible with acceptable accuracy, which the authors see as a basis for optimising post-sorting and dry-sort-re-dry decisions.

Portal interpretation (not claims of the paper)

The following is the portal's general reading, not something the study demonstrated.

• Mills that already weigh or scan boards at the green end could treat weight as a candidate input for a pre-kiln sort, and test internally whether it improves charge uniformity. The study shows a correlation on small laboratory batches, not a proven mill-scale benefit.
• A classifier is probably a better first project than a precise moisture predictor. An error that confuses "acceptable" with "slightly over-dried" matters less than a prediction that is wrong by several percentage points.
• If you log schedule and conditioning settings with each charge, you build the very dataset the paper shows to be useful. Good record-keeping is a low-cost starting point whatever modelling method follows.
• The aggressive Schedule II and the shorter Schedule I should not be read as recommendations; the paper compares them with a specific control for hemlock to a 12 percent target, and results will differ for other species, dimensions and kilns.

Limitations

The authors name several limitations themselves. Each run contained only 42 specimens, a small basis for a machine learning model. All drying targeted a single final moisture of 12 percent, so the model has not seen a wide range of targets. The regression overfitted, and test performance was poor. Only one algorithm (TreeNet gradient boosting) was tested; no comparison with neural networks, support vector machines or other methods was made. The study modelled differences between boards, not the moisture gradient inside a board (core versus shell), and did not model casehardening or drying defects such as checks and distortion. They suggest that acoustic, ultrasonic and near-infrared methods could provide additional inputs in future work.

Additional cautions from the portal: the material was 116 x 116 mm hemlock timber specimens 900 mm long in a 0.73 cubic metre laboratory kiln, which is far from a full-size industrial kiln with large airflow variation. Our reading of the published text also found some small internal inconsistencies, such as figure captions mentioning a different algorithm family than the methods section, and correlation coefficients given as percentages in places, so readers who want exact values should consult the original paper directly.

Conclusions

The study shows that final moisture content in a kiln charge depends on both the incoming wood and the way it was dried, and that a gradient boosting model benefits from knowing both. Initial weight stood out as the most useful and cheapest predictor. Predicting exact moisture content for single boards did not succeed on unseen data, but classifying boards into acceptable, over-dried and under-dried groups reached around 69 percent accuracy on the test portion, which the authors regard as workable for sorting strategies. For a sawmill, the main message is cautious: collect data on both board properties and kiln settings, think in terms of classes rather than exact numbers, and validate any model on your own species and kiln before relying on it.

Original research

• Authors: Sohrab Rahimi, Vahid Nasir, Stavros Avramidis, Farrokh Sassani
• Title: The Role of Drying Schedule and Conditioning in Moisture Uniformity in Wood: A Machine Learning Approach
• Journal/publisher: Polymers (MDPI), vol. 15, article 792
• Year: 2023
• DOI: 10.3390/polym15040792
• Link: https://doi.org/10.3390/polym15040792
• Licence: Creative Commons Attribution (CC BY 4.0)

This article is an independent summary written by the portal; the original paper should be consulted for full details, tables and the authors' own wording.