Mark Schubert
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mark Schubert has a total of 128 co-authors. Here are the top co-authors by number of shared publications:
- Francis Willis Mathew Robert Schwarze
Cellulose & Wood Materials Laboratory, Empa, Dübendorf, Switzerland
30 shared publications - M. J. Fuhr
Institute for Building Materials Computational Physics for Engineering Materials ETH Zurich Schafmattstrasse 6 HIF E18 CH-8093 Zurich Switzerland
6 shared publications S. Mourad
Process Technology Development, R+D, FISBA OPTIK AG, Rorschacherstrasse 268, 9016 St. Gallen, Switzerland
6 shared publicationsC. Stührk
5 shared publicationsSiegfried Fink
Albert-Ludwigs-Universität Freiburg, Professur für Forstbotanik, Bertoldstrasse 17, D-79085 Freiburg i. Br., Germany
4 shared publications- Greta Faccio
Empa – Swiss Federal Laboratories for Materials Science and Technology Laboratory for Biointerfaces Lerchenfeldstrasse 5 9014 St. Gallen Switzerland
3 shared publications
Latest Publications
- Open Access
- 18 April 2026
- Published by Springer Nature in Journal of Wood Science
- Vol. 72 (1)
Importance-driven bottleneck model: a multi-stage deep learning approach for analyzing swelling–shrinkage behavior of wood species and their mechanical properties
Experimental data were collected from defect-free specimens of 57 wood species in Central Europe, encompassing their structural, physical, chemical characteristics, and mechanical properties. The dataset highlights the complex relationships between physical properties (shrinkage–swelling behavior in three orthotropic directions), chemical composition (holocellulose, lignin, extractive content), structural characteristics (density, fiber length), and mechanical properties (compression, tensile and bending strength, hardness, and impact resistance), emphasizing the need for efficient and interpretable methods to identify their correlations. To address this issue, we developed an importance-driven bottleneck model, by revisiting the concept bottleneck model, which consists of a system of three connected neural networks; auxiliary, input, and output, that collaboratively work to investigate the relationship between the physical, structural, chemical, and mechanical features. The input and output networks are linked by a concept c bottleneck layer, with its features defined by analyzing wood’s structural characteristics and composition via feature importance analysis conducted in the auxiliary network. Density and fiber length were identified as the most important concept c features, yielding testing $$R^2$$ scores above 0.70 for predicting mechanical properties. Predictive accuracy increased to over 0.85 when a third component, either holocellulose or lignin content, was included in the bottleneck. By adding all five chemical and structural features used in concept c layer, the prediction accuracy was slightly increased. While the importance-driven bottleneck model with three or more concept c features was narrowly outperformed by an end-to-end network using shrinkage–swelling as direct inputs, it offered a superior balance between performance and interpretability. This mathematical strategy, forcing information through a concept bottleneck, creates an interpretable physics-inspired AI framework.- 18 April 2026
- Published by Springer Nature in Journal of Wood Science
- Vol. 72 (1)
Open AccessImportance-driven bottleneck model: a multi-stage deep learning approach for analyzing swelling–shrinkage behavior of wood species and their mechanical properties
Experimental data were collected from defect-free specimens of 57 wood species in Central Europe, encompassing their structural, physical, chemical characteristics, and mechanical properties. The dataset highlights the complex relationships between physical properties (shrinkage–swelling behavior in three orthotropic directions), chemical composition (holocellulose, lignin, extractive content), structural characteristics (density, fiber length), and mechanical properties (compression, tensile and bending strength, hardness, and impact resistance), emphasizing the need for efficient and interpretable methods to identify their correlations. To address this issue, we developed an importance-driven bottleneck model, by revisiting the concept bottleneck model, which consists of a system of three connected neural networks; auxiliary, input, and output, that collaboratively work to investigate the relationship between the physical, structural, chemical, and mechanical features. The input and output networks are linked by a concept c bottleneck layer, with its features defined by analyzing wood’s structural characteristics and composition via feature importance analysis conducted in the auxiliary network. Density and fiber length were identified as the most important concept c features, yielding testing $$R^2$$ scores above 0.70 for predicting mechanical properties. Predictive accuracy increased to over 0.85 when a third component, either holocellulose or lignin content, was included in the bottleneck. By adding all five chemical and structural features used in concept c layer, the prediction accuracy was slightly increased. While the importance-driven bottleneck model with three or more concept c features was narrowly outperformed by an end-to-end network using shrinkage–swelling as direct inputs, it offered a superior balance between performance and interpretability. This mathematical strategy, forcing information through a concept bottleneck, creates an interpretable physics-inspired AI framework. - Open Access
- 14 February 2026
- Published by Elsevier in Materials & Design
- Vol. 263
Multiscale descriptor-based framework for predicting on-demand debonding in silane-modified polyether interfaces
- 14 February 2026
- Published by Elsevier in Materials & Design
- Vol. 263
Open AccessMultiscale descriptor-based framework for predicting on-demand debonding in silane-modified polyether interfaces
- Open Access
- 13 January 2026
- Published by Springer Nature in Wood Science and Technology
- Vol. 60 (1) , 24
Bending performance improvement of spruce LVL using adhesive-bonded flax fiber composites
Increasing pressure on high-quality timber resources like Norway spruce necessitates enhanced material efficiency, driving interest in hybrid wood products. This study evaluated flax natural fiber-reinforced polymer reinforcement for spruce laminated veneer lumber. Reinforcement composites using conventional epoxy and bio-based polyfurfuryl alcohol matrices were bonded to the lumber with structural wood adhesives (melamine-urea-formaldehyde and polyurethane); direct reinforcement using these adhesives was also tested. Mechanical properties (tensile, shear, 4-point bending, impact) were assessed, including moisture-dependent behavior at 35–85% relative humidity. Results demonstrated effective bonding of prefabricated fiber composites using both wood adhesives, yielding significant increases in the lumber’s bending stiffness and strength. Notably, the polyfurfuryl-alcohol-based composites provided the highest stiffness gains, validating its bio-based potential, while epoxy-based composites exhibited superior toughness and impact resistance. Both systems maintained significant reinforcement despite expected moisture sensitivity. Direct reinforcement attempts using the wood adhesives revealed processing limitations. Findings confirm natural fiber reinforcements as a viable strategy for enhancing wood performance, enabling advanced hybrid wood products, contributing to more efficient use of wood resources.
- 13 January 2026
- Published by Springer Nature in Wood Science and Technology
- Vol. 60 (1) , 24
Open AccessBending performance improvement of spruce LVL using adhesive-bonded flax fiber composites
Increasing pressure on high-quality timber resources like Norway spruce necessitates enhanced material efficiency, driving interest in hybrid wood products. This study evaluated flax natural fiber-reinforced polymer reinforcement for spruce laminated veneer lumber. Reinforcement composites using conventional epoxy and bio-based polyfurfuryl alcohol matrices were bonded to the lumber with structural wood adhesives (melamine-urea-formaldehyde and polyurethane); direct reinforcement using these adhesives was also tested. Mechanical properties (tensile, shear, 4-point bending, impact) were assessed, including moisture-dependent behavior at 35–85% relative humidity. Results demonstrated effective bonding of prefabricated fiber composites using both wood adhesives, yielding significant increases in the lumber’s bending stiffness and strength. Notably, the polyfurfuryl-alcohol-based composites provided the highest stiffness gains, validating its bio-based potential, while epoxy-based composites exhibited superior toughness and impact resistance. Both systems maintained significant reinforcement despite expected moisture sensitivity. Direct reinforcement attempts using the wood adhesives revealed processing limitations. Findings confirm natural fiber reinforcements as a viable strategy for enhancing wood performance, enabling advanced hybrid wood products, contributing to more efficient use of wood resources.
- Open Access
- 24 May 2025
- Published by Springer Nature in Journal of Wood Science
- Vol. 71 (1) , 1-13
Optical detection of beetle-related indicators and stem quality in roundwood using convolutional neural networks
Accurate roundwood sorting is critical for improving resource efficiency in the timber industry. However, in many small sawmills, sorting is performed manually through visual inspection, often resulting in inconsistencies and misclassifications. Sorting wood based on macroscopic images using convolutional neural networks (CNN) is a cost-effective and efficient approach. However, current models focus primarily on cross-sectional images, limiting their ability to detect important features like knots and beetle infestations. We present an industrial dataset of 5,200 samples, each including both a trunk and a cross-sectional image. Using this dataset, we developed two models: a stem quality classification model and a beetle-indicator detection model. The first model uses only trunk images to extract stem information, supporting existing cross-sectional models. The beetle-indicator detection model utilizes both trunk and cross-sectional images, introducing a new detection capability to computer vision-based roundwood sorting. To ensure transparency and interpretability, we applied Explainable Artificial Intelligence (XAI) techniques to highlight key regions within the images that influenced the models’ predictions. The stem quality model achieves 80% accuracy in trunk feature analysis, and the beetle-indicator model reaches 89% accuracy. Together, these models improve automation based on convolutional neural networks by enabling detailed trunk analysis and early beetle infestation detection.
- 24 May 2025
- Published by Springer Nature in Journal of Wood Science
- Vol. 71 (1) , 1-13
Open AccessOptical detection of beetle-related indicators and stem quality in roundwood using convolutional neural networks
Accurate roundwood sorting is critical for improving resource efficiency in the timber industry. However, in many small sawmills, sorting is performed manually through visual inspection, often resulting in inconsistencies and misclassifications. Sorting wood based on macroscopic images using convolutional neural networks (CNN) is a cost-effective and efficient approach. However, current models focus primarily on cross-sectional images, limiting their ability to detect important features like knots and beetle infestations. We present an industrial dataset of 5,200 samples, each including both a trunk and a cross-sectional image. Using this dataset, we developed two models: a stem quality classification model and a beetle-indicator detection model. The first model uses only trunk images to extract stem information, supporting existing cross-sectional models. The beetle-indicator detection model utilizes both trunk and cross-sectional images, introducing a new detection capability to computer vision-based roundwood sorting. To ensure transparency and interpretability, we applied Explainable Artificial Intelligence (XAI) techniques to highlight key regions within the images that influenced the models’ predictions. The stem quality model achieves 80% accuracy in trunk feature analysis, and the beetle-indicator model reaches 89% accuracy. Together, these models improve automation based on convolutional neural networks by enabling detailed trunk analysis and early beetle infestation detection.
- Open Access
- 23 October 2024
- Published by Elsevier in International Journal of Adhesion and Adhesives
- Vol. 136
Bonding of beech wood to mortar with a novel epoxy hybrid-adhesive: Performance in dry and wet conditions
- 23 October 2024
- Published by Elsevier in International Journal of Adhesion and Adhesives
- Vol. 136
Open AccessBonding of beech wood to mortar with a novel epoxy hybrid-adhesive: Performance in dry and wet conditions
- Open Access
- 1 April 2024
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 21 (213) , 20230492
Probing the complexity of wood with computer vision: from pixels to properties
We use data produced by industrial wood grading machines to train a machine learning model for predicting strength-related properties of wood lamellae from colour images of their surfaces. The focus was on samples of Norway spruce (Picea abies) wood, which display visible fibre pattern formations on their surfaces. We used a pre-trained machine learning model based on the residual network ResNet50 that we trained with over 15 000 high-definition images labelled with the indicating properties measured by the grading machine. With the help of augmentation techniques, we were able to achieve a coefficient of determination (R2) value of just over 0.9. Considering the ever-increasing demand for construction-grade wood, we argue that computer vision should be considered a viable option for the automatic sorting and grading of wood lamellae in the future.
- 1 April 2024
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 21 (213) , 20230492
Open AccessProbing the complexity of wood with computer vision: from pixels to properties
We use data produced by industrial wood grading machines to train a machine learning model for predicting strength-related properties of wood lamellae from colour images of their surfaces. The focus was on samples of Norway spruce (Picea abies) wood, which display visible fibre pattern formations on their surfaces. We used a pre-trained machine learning model based on the residual network ResNet50 that we trained with over 15 000 high-definition images labelled with the indicating properties measured by the grading machine. With the help of augmentation techniques, we were able to achieve a coefficient of determination (R2) value of just over 0.9. Considering the ever-increasing demand for construction-grade wood, we argue that computer vision should be considered a viable option for the automatic sorting and grading of wood lamellae in the future.
- Open Access
- 4 January 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 246
Convolutional neural networks for quality and species sorting of roundwood with image and numerical data
- 4 January 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 246
Open AccessConvolutional neural networks for quality and species sorting of roundwood with image and numerical data
- Open Access
- 2 September 2022
- Published by Springer Nature in Wood Science and Technology
- Vol. 56 (5) , 1439-1453
Semi-supervised learning for quality control of high-value wood products
The quality control of wood products is often only checked at the end of the production process so that countermeasures can only be taken with a time delay in the event of fluctuations in product quality. This often leads to unnecessary and cost-intensive rejects. Furthermore, since quality control often requires additional procedural steps to be performed by a skilled worker, testing is time-consuming and costly. While traditional machine learning (ML) methods based on supervised learning have been used in the field with some success, the limited availability of labeled data is the major hurdle for further improving model performance. In the present study, the potential of enhancing the performance of the ML methods random forest (RF) and support vector machines (SVM) for quality classification by using semi-supervised learning (SSL) was investigated. Labeled and unlabeled data were provided by Swiss Wood Solutions AG, which produces densified wood for high-value wood products such as musical instruments. The developed approach includes labeling of the unlabeled data using SSL, training and 10k cross-validation of the ML algorithms RF and SVM, and determining the generalization ability using the hold-out test set. Based on the evaluation indices such as accuracy, F1-score, recall, false-positive-rate and confusion matrices, it was shown that SSL could enhance the prediction performance of the quality classification of ML models compared to the conventional supervised learning method. Despite having a small dataset, the work paves the way for future applications of SSL for wood quality assessment.
- 2 September 2022
- Published by Springer Nature in Wood Science and Technology
- Vol. 56 (5) , 1439-1453
Open AccessSemi-supervised learning for quality control of high-value wood products
The quality control of wood products is often only checked at the end of the production process so that countermeasures can only be taken with a time delay in the event of fluctuations in product quality. This often leads to unnecessary and cost-intensive rejects. Furthermore, since quality control often requires additional procedural steps to be performed by a skilled worker, testing is time-consuming and costly. While traditional machine learning (ML) methods based on supervised learning have been used in the field with some success, the limited availability of labeled data is the major hurdle for further improving model performance. In the present study, the potential of enhancing the performance of the ML methods random forest (RF) and support vector machines (SVM) for quality classification by using semi-supervised learning (SSL) was investigated. Labeled and unlabeled data were provided by Swiss Wood Solutions AG, which produces densified wood for high-value wood products such as musical instruments. The developed approach includes labeling of the unlabeled data using SSL, training and 10k cross-validation of the ML algorithms RF and SVM, and determining the generalization ability using the hold-out test set. Based on the evaluation indices such as accuracy, F1-score, recall, false-positive-rate and confusion matrices, it was shown that SSL could enhance the prediction performance of the quality classification of ML models compared to the conventional supervised learning method. Despite having a small dataset, the work paves the way for future applications of SSL for wood quality assessment.
- Open Access
- 1 September 2022
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 19 (194) , 20220349
Reconstructing radial stem size changes of trees with machine learning
Like many scientists, ecologists depend heavily on continuous uninterrupted data in order to understand better the object of their study. Although this might be straightforward to achieve under controlled laboratory conditions, the situation is easily complicated under field conditions where sensors and data transmission are affected by harsh weather, living organisms, changes in atmospheric conditions etc. This often results in parts of the data being corrupted or missing altogether. We propose the use of the most recent machine-learning techniques to reverse such data losses in multi-channel time series. In particular, we focus on tree stem growth data obtained from the TreeNet project, which monitors the changes in stem radius and environmental conditions of a few hundred trees across Switzerland. In the first part of the study, we test the performance of five architectures based on encoders and recurrent and convolutional neural networks, and we show that a deep neural network combining long short-term memory with one-dimensional convolutional layers performs the best. In the second part, we adopt this model to reconstruct the original TreeNet dataset, which we then use in a separate classification problem to show the effect of the proposed gap-filling procedure.
- 1 September 2022
- Published by The Royal Society in Journal of The Royal Society Interface
- Vol. 19 (194) , 20220349
Open AccessReconstructing radial stem size changes of trees with machine learning
Like many scientists, ecologists depend heavily on continuous uninterrupted data in order to understand better the object of their study. Although this might be straightforward to achieve under controlled laboratory conditions, the situation is easily complicated under field conditions where sensors and data transmission are affected by harsh weather, living organisms, changes in atmospheric conditions etc. This often results in parts of the data being corrupted or missing altogether. We propose the use of the most recent machine-learning techniques to reverse such data losses in multi-channel time series. In particular, we focus on tree stem growth data obtained from the TreeNet project, which monitors the changes in stem radius and environmental conditions of a few hundred trees across Switzerland. In the first part of the study, we test the performance of five architectures based on encoders and recurrent and convolutional neural networks, and we show that a deep neural network combining long short-term memory with one-dimensional convolutional layers performs the best. In the second part, we adopt this model to reconstruct the original TreeNet dataset, which we then use in a separate classification problem to show the effect of the proposed gap-filling procedure.
- Open Access
- 10 March 2022
- Published by Springer Nature in Research Square
Semi-Supervised Learning For Quality Control Of High-Value Wood Products
The quality control of wood products is often only checked at the end of the production process so that countermeasures can only be taken with a time delay in the event of fluctuations in product quality. This often leads to unnecessary and cost-intensive rejects. Furthermore, since quality control often requires additional procedural steps to be performed by a skilled worker, testing is time-consuming and costly. While traditional machine learning (ML) methods based on supervised learning have been used in the field with some success, the limited availability of labeled data is the major hurdle for further improving model performance. In the present study, the potential of enhancing the performance of the ML methods random forest (RF) and support vector machines (SVM) for quality classification by using semi-supervised learning (SSL) was investigated. Labelled and unlabeled data were provided by Swiss Wood Solutions AG, which produces densified wood for high-value wood products such as musical instruments. The developed approach includes labeling of the unlabeled data using SSL, training and 10k cross-validation of the ML algorithms RF and SVM, and determining the generalization ability using the hold-out test set. Based on the evaluation indices such as accuracy, F1-score, recall, false-positive-rate (FPR) and confusion matrices, it was shown that SSL could enhance the prediction performance of the quality classification of ML models compared to the conventional supervised learning method. Despite having a small data set, the work paves the way for future applications of semi-supervised ML for wood quality assessment.
- 10 March 2022
- Published by Springer Nature in Research Square
Open AccessSemi-Supervised Learning For Quality Control Of High-Value Wood Products
The quality control of wood products is often only checked at the end of the production process so that countermeasures can only be taken with a time delay in the event of fluctuations in product quality. This often leads to unnecessary and cost-intensive rejects. Furthermore, since quality control often requires additional procedural steps to be performed by a skilled worker, testing is time-consuming and costly. While traditional machine learning (ML) methods based on supervised learning have been used in the field with some success, the limited availability of labeled data is the major hurdle for further improving model performance. In the present study, the potential of enhancing the performance of the ML methods random forest (RF) and support vector machines (SVM) for quality classification by using semi-supervised learning (SSL) was investigated. Labelled and unlabeled data were provided by Swiss Wood Solutions AG, which produces densified wood for high-value wood products such as musical instruments. The developed approach includes labeling of the unlabeled data using SSL, training and 10k cross-validation of the ML algorithms RF and SVM, and determining the generalization ability using the hold-out test set. Based on the evaluation indices such as accuracy, F1-score, recall, false-positive-rate (FPR) and confusion matrices, it was shown that SSL could enhance the prediction performance of the quality classification of ML models compared to the conventional supervised learning method. Despite having a small data set, the work paves the way for future applications of semi-supervised ML for wood quality assessment.