Empa researcher Mark Schubert uses the enzyme laccase to alter wood properties, but the quest is quite challenging – a bit like attempting to locate the key to an unknown lock. Rather than using costly and time-consuming experiments, Schubert employs artificial intelligence as it allows him to reach the objective more quickly.
The enzyme laccase is able to alter the chemical structure of wood on its surface and thus facilitate biochemical modifications without changing the structure of the material. However, there are different laccases – and they don’t all work in every case. (Image credit: Thordis Rüggeberg)
The enzyme laccase is capable of changing the chemical structure of wood on its surface, thereby enabling biochemical changes without altering the material’s structure. For example, Empa scientists attached functional molecules to create antimicrobial or waterproof wood surfaces.
Adhesive wood fibers, which can be easily pressed to fiberboards without using any chemical binding agents, could also be developed. Such kinds of solvent-free fiberboards are utilized to insulate eco houses.
However, the problem is several variants of laccase exist, which vary in the design of the chemically active center; in addition, these variants do not react with the required substrate. Since it is very complicated to guess whether or not a specific laccase will react with a particular substrate, identifying appropriate pairs of laccase and substrate involves costly as well as time-consuming series of experiments. But this problem can be solved with molecular simulations: one merely requires an exact structural analysis of the laccase in order to mimic the mechanism of chemical reactions for each preferred combination on the computer. This however needs a high computer computing capacity and, despite this, would still be very costly and time-consuming
Conversely, “deep learning” provides a shortcut. Researchers train a computer program to detect patterns with data obtained from own experiments as well as from the literature: Which type of laccase oxidizes which substrate? What would be the optimized conditions for the required chemical process to occur? One best thing about this is that even if all the details about the chemical mechanism are not known, the search will still work.
Major progress in the last seven years
Key to this success is the availability of the data in an appropriate form as well as the design of the deep-learning network. Schubert, in fact, has already been working with neuronal networks for more than seven years. His initial project on the topic comes from 2012, the newest from 2018. “In the past, we worked with shallow neuronal networks: an input layer, a hidden layer and an output layer. Today, we work with considerably more complex networks. They contain several hidden layers and are so much more powerful.”
Using known datasets, Schubert trains his algorithms and then tests them with datasets that have never been seen by the system. His reports on the sturdiness of his “smart wood search engine” are incredible. Earlier, Schubert can only use meaningful data that have been carefully chosen to obtain decent outcomes. In the meantime, he is also using partly unusable data piles to test his systems. The machine can identify what it can use and what it cannot use.
Industrial application of KI
Thanks to the robustness of the system, the deep-learning machine can be utilized by industry. At Pavatex, the partner company which Schubert has been associating with for some time, produces self-adhesive insulation boards. The manufacturing process is full of sensors; vast amounts of data build up that reveal “something” regarding the quality of the manufactured boards. But what? The smart wood search engine of Schubert identifies the connection.
At present, Schubert is working on improving the production in this manner. In case something goes wrong at one point during the fiber processing process, the production can be modified before the end product’s quality is impacted. This will not only save expensive checks on the final product but would also reduce the error rate in the manufacturing process.