How are engineered wood producers using production data?

Inside the plant, one of the most useful jobs for production data is predicting panel properties from process data before the lab test comes back. This is about a mill's own process records, not the industry output statistics that trade bodies publish. Panel plants record resin dosing, wood moisture, press temperature, pressure and time, and line speed continuously. Their key quality tests, such as internal bond, are destructive and run on samples hours later. Models trained on the two together can predict the property as the board is made. The published evidence is on particleboard; we expect the approach to carry over to OSB and other panels with their own models.

What the usual answer says

Search results read the question as industry production statistics or product carbon footprints. Guides to AI in wood products list process control, quality monitoring, predictive maintenance, resin optimisation and defect detection. None says which data drives them, or what makes a prediction safe to act on.

What the research shows

Two studies used data from particleboard production lines, and we read both as abstracts. Beilong Zhang and colleagues, publishing in the Journal of Wood Science in 2022, trained a model on 724 sets of gluing parameters and internal bond results from a particleboard line and tested it on 181 more. They concluded it "can be used to real-time predict the IB in the PB production line". They describe the problem plainly: "The particleboard (PB) production is an extremely complex process, many operating parameters affecting panel quality".

Francisco García Fernández and colleagues, in Applied Sciences in April 2025, trained neural networks on "experimental data taken from real production processes" and used bootstrapping to give each prediction a confidence interval. The models reached a determination coefficient of 0.96, and the intervals covered the true value 93% of the time. Their conclusion is the one that matters for a plant manager: the method "allows decision making based on confidence intervals rather than individual values".

Why the range matters

A single predicted value invites trouble. If a model predicts internal bond comfortably above specification, a plant might cut resin. If the prediction was wrong, the board fails, and the lab finds out hours and many panels later. A prediction with a range says how far above specification the board is, with what confidence.

That is the argument for margins, and it is ours: neither study reports a resin saving. If a plant sets resin with a margin for not knowing, a range that holds up against the lab is what would let it narrow that margin, only while the lower end stays above specification. The 93% coverage also means about 7 predictions in 100 fall outside their range. At a structural plant, a lower-bound miss is a failed lot, and emission limits and certification testing set their own floor. So any narrowing is small, gradual and checked.

What data it takes

Process records with accurate timestamps: resin and wax dosing, furnish moisture, mat weight, press temperatures, pressures and times, and line speed. Lab results with the time the tested board was pressed, not the time it was tested. And a way to join the two, board by board or at least minute by minute. What we typically find is process data in the control system and lab results in a separate database or spreadsheet, with no shared key but time.

How to start

Pick one property and one line. Gather several months of process records and lab results, and join them by the time each tested board was pressed. Train a model that gives a range, and run it alongside the lab for a few months without acting on it. When the lab result falls inside the predicted range as often as the model claims, the model has earned trust. Then narrow one setting a little, and keep the lab testing as before.

Where else the joined data pays

Once process and quality data are joined, other comparisons become routine: resin use by furnish source or species mix, and press cycle length against quality outcome.

When it doesn't apply

Plants that don't store process data with timestamps, or keep lab results only on paper, need to fix that first. Very small plants with few quality tests may not have enough history to train a model. And OSB, plywood and structural products such as LVL and CLT have their own tests and processes, where the approach carries over but the particleboard models do not.

Quarri for wood products is built around how a wood products plant runs, where the order book meets real capacity.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.