Is a digital twin worth building for a forest or a mill?

Rarely as a whole, for a mill; more often for a forest. A digital twin of a forest or a mill is a continuously updated replica of the physical system, fed by sensors, inventory and machine data. In the one published cost claim we traced, the gain credited to a twin came from one part inside it, and a single decision model can be built and measured on its own. The better question is which decision you would simulate most often and what it costs when it is wrong. In a mill, build a model of that decision first. In a forest, the shared inventory a twin needs may be the first thing to build anyway.

What the usual answer says

The descriptions that rank for this question are enthusiastic. An Oregon State University article from July 2026 describes digital twin forests as "Continuously updated virtual replicas that mirror changing conditions in real forests," combining satellite imagery, LiDAR, field measurements and sensor networks. An editorial in Frontiers in Plant Science by Lv and colleagues puts it more briefly: "Digital twins reflect the full life cycle of objects by mapping them into a virtual space."

Neither says what a twin costs to keep current, or which of its benefits could be had without it.

Where the gains actually come from

The example is forest restoration planting, because it is the only twin cost claim we found with its evidence published. A 2025 review in Future Internet by Nophea Sasaki and Issei Abe proposes a four-layer digital twin for restoration. Its headline evidence is cost. Citing a 2022 white paper by Oakes and colleagues, it says traditional programmes "typically report per-tree planting costs ranging from USD 2.00 to 3.75". Systems of the kind it describes have reached costs, it says, "as low as USD 0.11 per tree, with field deployment speeds up to 25 times faster than manual approaches".

The companies it cites for those figures plant from the air. Pilots from two of them, it says, showed that "drone-based reforestation can reduce planting costs by over 80%". The saving comes from seeding by drone instead of by hand. The review credits the whole architecture, including a data layer, an analytics layer and dashboards, with a result that the drones produced. The same paper notes that for one aerial planter "seedling survival rates can be below 20%". A cost per seed pod dropped is not a cost per tree established.

The white paper the review cites for its figures, from Trillion Trees, says something different. Its $2 and $3.75 a tree are two project examples, and it contains no drone figures at all. Its warning runs the other way. Organisations marketing low per-tree prices, it says, often need extra grants, "meaning the true costs exceed those used in promotions". A twin's headline gain, in this case, traces to one component, and its baseline traces to a source that doesn't say it.

A model of one decision

A single-decision model can be measured on its own. From Quarri's own work with a sawmill: a production scheduler re-solves an order book in about a second, and lands within 1 to 1.6% of the best possible schedule found by exhaustive search. That is a model of one decision, which orders to cut in which sequence, measured against a known best. It does not need a three-dimensional replica of the mill.

A full twin could contain that scheduler. It would add cost and upkeep for everything else it mirrors, and the value of the scheduler would be the same.

Forests are different

In a forest, the case for a shared model is stronger. The Oregon State article's point is that twins replace "inventories collected every several years" with continuous monitoring. Harvest sequencing, fire risk and carbon all draw on that same inventory, so a model of any one of them needs most of the twin's data first. For a forest manager, we would expect several decisions near the top of the list to share one inventory.

How to decide

List the decisions you would want to simulate. For a mill that might be the weekly schedule, a change to the log sort, a new machine. For a forest it might be harvest sequencing, road building, or planting. Next to each, write how often it is made and what a wrong call costs.

Build a model of the decision at the top of that list, on the data it actually needs, and measure it against a known answer. If several decisions near the top draw on the same data, a shared model of the operation starts to make sense. That is when a twin is worth considering. If they do not, separate decision models will do more for less.

When it doesn't apply

Large capital changes, such as rebuilding a mill line or adding a paper machine, are often studied with detailed engineering simulations, and for one-off decisions of that size a full model can pay for itself once. Research programmes that study forests as systems, rather than making operational decisions, have reasons to build complete replicas. And any twin is only as current as the data that feeds it.

Quarri for sawmills is built around how a sawmill runs, from log intake to shipped order.

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.